Search Results

Search found 4132 results on 166 pages for 'collin white'.

Page 15/166 | < Previous Page | 11 12 13 14 15 16 17 18 19 20 21 22  | Next Page >

  • Natural Search Engine Optimization - Don't "Game the System" Or You Will Get Banned!

    When focusing on natural search engine optimization, it is important that you keep the process "white hat." You see, when it comes to SEO, there are basically three schools of thought: White hat, Gray hat, and Black hat. As you can probably infer, white hat is following the rules, gray hat is a little in between, and black hat is going against parameters that Google and other major search engines have set for ethical SEO practices.

    Read the article

  • Why is 1px sometimes 2px when specified in Android XML?

    - by Daniel Lew
    I've got a desire for a one-pixel divider line, just for looks. I thought I could accomplish this using a View of height 1px, with a defined background. However, I'm getting some very odd behavior on different devices - sometimes the 1px ends up as 2px. Take this sample layout for example: <LinearLayout xmlns:android="http://schemas.android.com/apk/res/android" android:orientation="vertical" android:layout_width="fill_parent" android:layout_height="fill_parent"> <View android:layout_width="fill_parent" android:layout_height="1px" android:background="@android:color/white" android:layout_marginBottom="4dp" /> <View android:layout_width="fill_parent" android:layout_height="1px" android:background="@android:color/white" android:layout_marginBottom="4dp" /> <View android:layout_width="fill_parent" android:layout_height="1px" android:background="@android:color/white" android:layout_marginBottom="4dp" /> <View android:layout_width="fill_parent" android:layout_height="1px" android:background="@android:color/white" android:layout_marginBottom="4dp" /> <View android:layout_width="fill_parent" android:layout_height="1px" android:background="@android:color/white" android:layout_marginBottom="4dp" /> <View android:layout_width="fill_parent" android:layout_height="1px" android:background="@android:color/white" android:layout_marginBottom="4dp" /> <View android:layout_width="fill_parent" android:layout_height="1px" android:background="@android:color/white" android:layout_marginBottom="4dp" /> </LinearLayout> When run on my G1, this comes out fine. But on the Nexus One, it alternates between 1px lines and 2px lines. Does anyone know where this is going awry? Why does Android sometimes make 1px into 2px?

    Read the article

  • Toorcon 15 (2013)

    - by danx
    The Toorcon gang (senior staff): h1kari (founder), nfiltr8, and Geo Introduction to Toorcon 15 (2013) A Tale of One Software Bypass of MS Windows 8 Secure Boot Breaching SSL, One Byte at a Time Running at 99%: Surviving an Application DoS Security Response in the Age of Mass Customized Attacks x86 Rewriting: Defeating RoP and other Shinanighans Clowntown Express: interesting bugs and running a bug bounty program Active Fingerprinting of Encrypted VPNs Making Attacks Go Backwards Mask Your Checksums—The Gorry Details Adventures with weird machines thirty years after "Reflections on Trusting Trust" Introduction to Toorcon 15 (2013) Toorcon 15 is the 15th annual security conference held in San Diego. I've attended about a third of them and blogged about previous conferences I attended here starting in 2003. As always, I've only summarized the talks I attended and interested me enough to write about them. Be aware that I may have misrepresented the speaker's remarks and that they are not my remarks or opinion, or those of my employer, so don't quote me or them. Those seeking further details may contact the speakers directly or use The Google. For some talks, I have a URL for further information. A Tale of One Software Bypass of MS Windows 8 Secure Boot Andrew Furtak and Oleksandr Bazhaniuk Yuri Bulygin, Oleksandr ("Alex") Bazhaniuk, and (not present) Andrew Furtak Yuri and Alex talked about UEFI and Bootkits and bypassing MS Windows 8 Secure Boot, with vendor recommendations. They previously gave this talk at the BlackHat 2013 conference. MS Windows 8 Secure Boot Overview UEFI (Unified Extensible Firmware Interface) is interface between hardware and OS. UEFI is processor and architecture independent. Malware can replace bootloader (bootx64.efi, bootmgfw.efi). Once replaced can modify kernel. Trivial to replace bootloader. Today many legacy bootkits—UEFI replaces them most of them. MS Windows 8 Secure Boot verifies everything you load, either through signatures or hashes. UEFI firmware relies on secure update (with signed update). You would think Secure Boot would rely on ROM (such as used for phones0, but you can't do that for PCs—PCs use writable memory with signatures DXE core verifies the UEFI boat loader(s) OS Loader (winload.efi, winresume.efi) verifies the OS kernel A chain of trust is established with a root key (Platform Key, PK), which is a cert belonging to the platform vendor. Key Exchange Keys (KEKs) verify an "authorized" database (db), and "forbidden" database (dbx). X.509 certs with SHA-1/SHA-256 hashes. Keys are stored in non-volatile (NV) flash-based NVRAM. Boot Services (BS) allow adding/deleting keys (can't be accessed once OS starts—which uses Run-Time (RT)). Root cert uses RSA-2048 public keys and PKCS#7 format signatures. SecureBoot — enable disable image signature checks SetupMode — update keys, self-signed keys, and secure boot variables CustomMode — allows updating keys Secure Boot policy settings are: always execute, never execute, allow execute on security violation, defer execute on security violation, deny execute on security violation, query user on security violation Attacking MS Windows 8 Secure Boot Secure Boot does NOT protect from physical access. Can disable from console. Each BIOS vendor implements Secure Boot differently. There are several platform and BIOS vendors. It becomes a "zoo" of implementations—which can be taken advantage of. Secure Boot is secure only when all vendors implement it correctly. Allow only UEFI firmware signed updates protect UEFI firmware from direct modification in flash memory protect FW update components program SPI controller securely protect secure boot policy settings in nvram protect runtime api disable compatibility support module which allows unsigned legacy Can corrupt the Platform Key (PK) EFI root certificate variable in SPI flash. If PK is not found, FW enters setup mode wich secure boot turned off. Can also exploit TPM in a similar manner. One is not supposed to be able to directly modify the PK in SPI flash from the OS though. But they found a bug that they can exploit from User Mode (undisclosed) and demoed the exploit. It loaded and ran their own bootkit. The exploit requires a reboot. Multiple vendors are vulnerable. They will disclose this exploit to vendors in the future. Recommendations: allow only signed updates protect UEFI fw in ROM protect EFI variable store in ROM Breaching SSL, One Byte at a Time Yoel Gluck and Angelo Prado Angelo Prado and Yoel Gluck, Salesforce.com CRIME is software that performs a "compression oracle attack." This is possible because the SSL protocol doesn't hide length, and because SSL compresses the header. CRIME requests with every possible character and measures the ciphertext length. Look for the plaintext which compresses the most and looks for the cookie one byte-at-a-time. SSL Compression uses LZ77 to reduce redundancy. Huffman coding replaces common byte sequences with shorter codes. US CERT thinks the SSL compression problem is fixed, but it isn't. They convinced CERT that it wasn't fixed and they issued a CVE. BREACH, breachattrack.com BREACH exploits the SSL response body (Accept-Encoding response, Content-Encoding). It takes advantage of the fact that the response is not compressed. BREACH uses gzip and needs fairly "stable" pages that are static for ~30 seconds. It needs attacker-supplied content (say from a web form or added to a URL parameter). BREACH listens to a session's requests and responses, then inserts extra requests and responses. Eventually, BREACH guesses a session's secret key. Can use compression to guess contents one byte at-a-time. For example, "Supersecret SupersecreX" (a wrong guess) compresses 10 bytes, and "Supersecret Supersecret" (a correct guess) compresses 11 bytes, so it can find each character by guessing every character. To start the guess, BREACH needs at least three known initial characters in the response sequence. Compression length then "leaks" information. Some roadblocks include no winners (all guesses wrong) or too many winners (multiple possibilities that compress the same). The solutions include: lookahead (guess 2 or 3 characters at-a-time instead of 1 character). Expensive rollback to last known conflict check compression ratio can brute-force first 3 "bootstrap" characters, if needed (expensive) block ciphers hide exact plain text length. Solution is to align response in advance to block size Mitigations length: use variable padding secrets: dynamic CSRF tokens per request secret: change over time separate secret to input-less servlets Future work eiter understand DEFLATE/GZIP HTTPS extensions Running at 99%: Surviving an Application DoS Ryan Huber Ryan Huber, Risk I/O Ryan first discussed various ways to do a denial of service (DoS) attack against web services. One usual method is to find a slow web page and do several wgets. Or download large files. Apache is not well suited at handling a large number of connections, but one can put something in front of it Can use Apache alternatives, such as nginx How to identify malicious hosts short, sudden web requests user-agent is obvious (curl, python) same url requested repeatedly no web page referer (not normal) hidden links. hide a link and see if a bot gets it restricted access if not your geo IP (unless the website is global) missing common headers in request regular timing first seen IP at beginning of attack count requests per hosts (usually a very large number) Use of captcha can mitigate attacks, but you'll lose a lot of genuine users. Bouncer, goo.gl/c2vyEc and www.github.com/rawdigits/Bouncer Bouncer is software written by Ryan in netflow. Bouncer has a small, unobtrusive footprint and detects DoS attempts. It closes blacklisted sockets immediately (not nice about it, no proper close connection). Aggregator collects requests and controls your web proxies. Need NTP on the front end web servers for clean data for use by bouncer. Bouncer is also useful for a popularity storm ("Slashdotting") and scraper storms. Future features: gzip collection data, documentation, consumer library, multitask, logging destroyed connections. Takeaways: DoS mitigation is easier with a complete picture Bouncer designed to make it easier to detect and defend DoS—not a complete cure Security Response in the Age of Mass Customized Attacks Peleus Uhley and Karthik Raman Peleus Uhley and Karthik Raman, Adobe ASSET, blogs.adobe.com/asset/ Peleus and Karthik talked about response to mass-customized exploits. Attackers behave much like a business. "Mass customization" refers to concept discussed in the book Future Perfect by Stan Davis of Harvard Business School. Mass customization is differentiating a product for an individual customer, but at a mass production price. For example, the same individual with a debit card receives basically the same customized ATM experience around the world. Or designing your own PC from commodity parts. Exploit kits are another example of mass customization. The kits support multiple browsers and plugins, allows new modules. Exploit kits are cheap and customizable. Organized gangs use exploit kits. A group at Berkeley looked at 77,000 malicious websites (Grier et al., "Manufacturing Compromise: The Emergence of Exploit-as-a-Service", 2012). They found 10,000 distinct binaries among them, but derived from only a dozen or so exploit kits. Characteristics of Mass Malware: potent, resilient, relatively low cost Technical characteristics: multiple OS, multipe payloads, multiple scenarios, multiple languages, obfuscation Response time for 0-day exploits has gone down from ~40 days 5 years ago to about ~10 days now. So the drive with malware is towards mass customized exploits, to avoid detection There's plenty of evicence that exploit development has Project Manager bureaucracy. They infer from the malware edicts to: support all versions of reader support all versions of windows support all versions of flash support all browsers write large complex, difficult to main code (8750 lines of JavaScript for example Exploits have "loose coupling" of multipe versions of software (adobe), OS, and browser. This allows specific attacks against specific versions of multiple pieces of software. Also allows exploits of more obscure software/OS/browsers and obscure versions. Gave examples of exploits that exploited 2, 3, 6, or 14 separate bugs. However, these complete exploits are more likely to be buggy or fragile in themselves and easier to defeat. Future research includes normalizing malware and Javascript. Conclusion: The coming trend is that mass-malware with mass zero-day attacks will result in mass customization of attacks. x86 Rewriting: Defeating RoP and other Shinanighans Richard Wartell Richard Wartell The attack vector we are addressing here is: First some malware causes a buffer overflow. The malware has no program access, but input access and buffer overflow code onto stack Later the stack became non-executable. The workaround malware used was to write a bogus return address to the stack jumping to malware Later came ASLR (Address Space Layout Randomization) to randomize memory layout and make addresses non-deterministic. The workaround malware used was to jump t existing code segments in the program that can be used in bad ways "RoP" is Return-oriented Programming attacks. RoP attacks use your own code and write return address on stack to (existing) expoitable code found in program ("gadgets"). Pinkie Pie was paid $60K last year for a RoP attack. One solution is using anti-RoP compilers that compile source code with NO return instructions. ASLR does not randomize address space, just "gadgets". IPR/ILR ("Instruction Location Randomization") randomizes each instruction with a virtual machine. Richard's goal was to randomize a binary with no source code access. He created "STIR" (Self-Transofrming Instruction Relocation). STIR disassembles binary and operates on "basic blocks" of code. The STIR disassembler is conservative in what to disassemble. Each basic block is moved to a random location in memory. Next, STIR writes new code sections with copies of "basic blocks" of code in randomized locations. The old code is copied and rewritten with jumps to new code. the original code sections in the file is marked non-executible. STIR has better entropy than ASLR in location of code. Makes brute force attacks much harder. STIR runs on MS Windows (PEM) and Linux (ELF). It eliminated 99.96% or more "gadgets" (i.e., moved the address). Overhead usually 5-10% on MS Windows, about 1.5-4% on Linux (but some code actually runs faster!). The unique thing about STIR is it requires no source access and the modified binary fully works! Current work is to rewrite code to enforce security policies. For example, don't create a *.{exe,msi,bat} file. Or don't connect to the network after reading from the disk. Clowntown Express: interesting bugs and running a bug bounty program Collin Greene Collin Greene, Facebook Collin talked about Facebook's bug bounty program. Background at FB: FB has good security frameworks, such as security teams, external audits, and cc'ing on diffs. But there's lots of "deep, dark, forgotten" parts of legacy FB code. Collin gave several examples of bountied bugs. Some bounty submissions were on software purchased from a third-party (but bounty claimers don't know and don't care). We use security questions, as does everyone else, but they are basically insecure (often easily discoverable). Collin didn't expect many bugs from the bounty program, but they ended getting 20+ good bugs in first 24 hours and good submissions continue to come in. Bug bounties bring people in with different perspectives, and are paid only for success. Bug bounty is a better use of a fixed amount of time and money versus just code review or static code analysis. The Bounty program started July 2011 and paid out $1.5 million to date. 14% of the submissions have been high priority problems that needed to be fixed immediately. The best bugs come from a small % of submitters (as with everything else)—the top paid submitters are paid 6 figures a year. Spammers like to backstab competitors. The youngest sumitter was 13. Some submitters have been hired. Bug bounties also allows to see bugs that were missed by tools or reviews, allowing improvement in the process. Bug bounties might not work for traditional software companies where the product has release cycle or is not on Internet. Active Fingerprinting of Encrypted VPNs Anna Shubina Anna Shubina, Dartmouth Institute for Security, Technology, and Society (I missed the start of her talk because another track went overtime. But I have the DVD of the talk, so I'll expand later) IPsec leaves fingerprints. Using netcat, one can easily visually distinguish various crypto chaining modes just from packet timing on a chart (example, DES-CBC versus AES-CBC) One can tell a lot about VPNs just from ping roundtrips (such as what router is used) Delayed packets are not informative about a network, especially if far away from the network More needed to explore about how TCP works in real life with respect to timing Making Attacks Go Backwards Fuzzynop FuzzyNop, Mandiant This talk is not about threat attribution (finding who), product solutions, politics, or sales pitches. But who are making these malware threats? It's not a single person or group—they have diverse skill levels. There's a lot of fat-fingered fumblers out there. Always look for low-hanging fruit first: "hiding" malware in the temp, recycle, or root directories creation of unnamed scheduled tasks obvious names of files and syscalls ("ClearEventLog") uncleared event logs. Clearing event log in itself, and time of clearing, is a red flag and good first clue to look for on a suspect system Reverse engineering is hard. Disassembler use takes practice and skill. A popular tool is IDA Pro, but it takes multiple interactive iterations to get a clean disassembly. Key loggers are used a lot in targeted attacks. They are typically custom code or built in a backdoor. A big tip-off is that non-printable characters need to be printed out (such as "[Ctrl]" "[RightShift]") or time stamp printf strings. Look for these in files. Presence is not proof they are used. Absence is not proof they are not used. Java exploits. Can parse jar file with idxparser.py and decomile Java file. Java typially used to target tech companies. Backdoors are the main persistence mechanism (provided externally) for malware. Also malware typically needs command and control. Application of Artificial Intelligence in Ad-Hoc Static Code Analysis John Ashaman John Ashaman, Security Innovation Initially John tried to analyze open source files with open source static analysis tools, but these showed thousands of false positives. Also tried using grep, but tis fails to find anything even mildly complex. So next John decided to write his own tool. His approach was to first generate a call graph then analyze the graph. However, the problem is that making a call graph is really hard. For example, one problem is "evil" coding techniques, such as passing function pointer. First the tool generated an Abstract Syntax Tree (AST) with the nodes created from method declarations and edges created from method use. Then the tool generated a control flow graph with the goal to find a path through the AST (a maze) from source to sink. The algorithm is to look at adjacent nodes to see if any are "scary" (a vulnerability), using heuristics for search order. The tool, called "Scat" (Static Code Analysis Tool), currently looks for C# vulnerabilities and some simple PHP. Later, he plans to add more PHP, then JSP and Java. For more information see his posts in Security Innovation blog and NRefactory on GitHub. Mask Your Checksums—The Gorry Details Eric (XlogicX) Davisson Eric (XlogicX) Davisson Sometimes in emailing or posting TCP/IP packets to analyze problems, you may want to mask the IP address. But to do this correctly, you need to mask the checksum too, or you'll leak information about the IP. Problem reports found in stackoverflow.com, sans.org, and pastebin.org are usually not masked, but a few companies do care. If only the IP is masked, the IP may be guessed from checksum (that is, it leaks data). Other parts of packet may leak more data about the IP. TCP and IP checksums both refer to the same data, so can get more bits of information out of using both checksums than just using one checksum. Also, one can usually determine the OS from the TTL field and ports in a packet header. If we get hundreds of possible results (16x each masked nibble that is unknown), one can do other things to narrow the results, such as look at packet contents for domain or geo information. With hundreds of results, can import as CSV format into a spreadsheet. Can corelate with geo data and see where each possibility is located. Eric then demoed a real email report with a masked IP packet attached. Was able to find the exact IP address, given the geo and university of the sender. Point is if you're going to mask a packet, do it right. Eric wouldn't usually bother, but do it correctly if at all, to not create a false impression of security. Adventures with weird machines thirty years after "Reflections on Trusting Trust" Sergey Bratus Sergey Bratus, Dartmouth College (and Julian Bangert and Rebecca Shapiro, not present) "Reflections on Trusting Trust" refers to Ken Thompson's classic 1984 paper. "You can't trust code that you did not totally create yourself." There's invisible links in the chain-of-trust, such as "well-installed microcode bugs" or in the compiler, and other planted bugs. Thompson showed how a compiler can introduce and propagate bugs in unmodified source. But suppose if there's no bugs and you trust the author, can you trust the code? Hell No! There's too many factors—it's Babylonian in nature. Why not? Well, Input is not well-defined/recognized (code's assumptions about "checked" input will be violated (bug/vunerabiliy). For example, HTML is recursive, but Regex checking is not recursive. Input well-formed but so complex there's no telling what it does For example, ELF file parsing is complex and has multiple ways of parsing. Input is seen differently by different pieces of program or toolchain Any Input is a program input executes on input handlers (drives state changes & transitions) only a well-defined execution model can be trusted (regex/DFA, PDA, CFG) Input handler either is a "recognizer" for the inputs as a well-defined language (see langsec.org) or it's a "virtual machine" for inputs to drive into pwn-age ELF ABI (UNIX/Linux executible file format) case study. Problems can arise from these steps (without planting bugs): compiler linker loader ld.so/rtld relocator DWARF (debugger info) exceptions The problem is you can't really automatically analyze code (it's the "halting problem" and undecidable). Only solution is to freeze code and sign it. But you can't freeze everything! Can't freeze ASLR or loading—must have tables and metadata. Any sufficiently complex input data is the same as VM byte code Example, ELF relocation entries + dynamic symbols == a Turing Complete Machine (TM). @bxsays created a Turing machine in Linux from relocation data (not code) in an ELF file. For more information, see Rebecca "bx" Shapiro's presentation from last year's Toorcon, "Programming Weird Machines with ELF Metadata" @bxsays did same thing with Mach-O bytecode Or a DWARF exception handling data .eh_frame + glibc == Turning Machine X86 MMU (IDT, GDT, TSS): used address translation to create a Turning Machine. Page handler reads and writes (on page fault) memory. Uses a page table, which can be used as Turning Machine byte code. Example on Github using this TM that will fly a glider across the screen Next Sergey talked about "Parser Differentials". That having one input format, but two parsers, will create confusion and opportunity for exploitation. For example, CSRs are parsed during creation by cert requestor and again by another parser at the CA. Another example is ELF—several parsers in OS tool chain, which are all different. Can have two different Program Headers (PHDRs) because ld.so parses multiple PHDRs. The second PHDR can completely transform the executable. This is described in paper in the first issue of International Journal of PoC. Conclusions trusting computers not only about bugs! Bugs are part of a problem, but no by far all of it complex data formats means bugs no "chain of trust" in Babylon! (that is, with parser differentials) we need to squeeze complexity out of data until data stops being "code equivalent" Further information See and langsec.org. USENIX WOOT 2013 (Workshop on Offensive Technologies) for "weird machines" papers and videos.

    Read the article

  • Shortest Path algorithm of a different kind

    - by Ram Bhat
    Hey guys, Lets say you have a grid like this (made randomly) Now lets say you have a car starting randomly from one of the while boxes, what would be the shortest path to go through each one of the white boxes? you can visit each white box as many times as you want and cant Jump over the black boxes. The black boxes are like walls. In simple words you can move from white box to white box only.. You can move in any direction, even diagonally.

    Read the article

  • Ajax Tabs implementation problem .

    - by SmartDev
    Hi , I have Implement ajax tabs and i have four tabs in it . In the four tabs i have four grid views with paging and sorting .The tabs are looking good i can see the grid ,but the problem is my first tab sorting works fine, where if i click on any other tab and click on the grid it goes to my first tab again . One more thing i want to change the background color of each tab. Can anyone help please here is my source code: <asp:Content ID="Content1" ContentPlaceHolderID="ContentPlaceHolder1" Runat="Server"> <asp:ScriptManager ID="ScMMyTabs" runat="server"> </asp:ScriptManager> <cc1:TabContainer ID="TCMytabs" ActiveTabIndex="0" runat="server"> <cc1:TabPanel ID="TpMyreq" runat="server" CssClass="TabBackground" HeaderText="My request"> <ContentTemplate> <table> <tr> <td> <asp:Button ID="btnexportMyRequestCsu" runat="server" Text="Export To Excel" CssClass="LabelDisplay" OnClick="btnexportMyRequestCsu_Click" /> </td> </tr> <tr> <td> <asp:GridView ID="GdvMyrequest" runat="server" CssClass="Mytabs" BackColor="White" BorderColor="White" BorderStyle="Ridge" BorderWidth="2px" CellPadding="3" CellSpacing="1" GridLines="None" OnPageIndexChanging="GdvMyrequest_PageIndexChanging" OnSorting="GdvMyrequest_Sorting" EmptyDataText="No request found for this user"> <RowStyle BackColor="#DEDFDE" ForeColor="Black" /> <FooterStyle BackColor="#C6C3C6" ForeColor="Black" /> <PagerStyle BackColor="Control" ForeColor="Gray" HorizontalAlign="Left" /> <SelectedRowStyle BackColor="#9471DE" Font-Bold="True" ForeColor="White" /> <HeaderStyle BackColor="#4A3C8C" Font-Bold="True" ForeColor="#E7E7FF" /> <PagerSettings Position="TopAndBottom" /> <Columns> <asp:TemplateField> <HeaderTemplate> Row No </HeaderTemplate> <ItemTemplate> <%# Container.DataItemIndex + 1 %> </ItemTemplate> </asp:TemplateField> </Columns> </asp:GridView> </td> </tr> <tr> <td> <asp:Label ID="lblmessmyrequestAhk" runat="server" CssClass="labelmess"></asp:Label> </td> </tr> </table> </ContentTemplate> </cc1:TabPanel> <cc1:TabPanel ID="TpMyPaymentCc" runat="server" HeaderText="Payments Credit Card" > <ContentTemplate> <table> <tr> <td> <asp:GridView ID="GdvmypaymentsCc" runat="server" CssClass="Mytabs" BackColor="White" BorderColor="White" BorderStyle="Ridge" BorderWidth="2px" CellPadding="3" CellSpacing="1" GridLines="None" OnPageIndexChanging="GdvmypaymentsCc_PageIndexChanging" OnSorting="GdvmypaymentsCc_Sorting" EmptyDataText="No Data"> <RowStyle BackColor="#DEDFDE" ForeColor="Black" /> <FooterStyle BackColor="#C6C3C6" ForeColor="Black" /> <PagerStyle BackColor="Control" ForeColor="Gray" HorizontalAlign="Left" /> <SelectedRowStyle BackColor="#9471DE" Font-Bold="True" ForeColor="White" /> <HeaderStyle BackColor="#4A3C8C" Font-Bold="True" ForeColor="#E7E7FF" /> <PagerSettings Position="TopAndBottom" /> </asp:GridView> </td> </tr> <tr> <td> <asp:Label ID="lblmessmypaymentsCsu" runat="server" CssClass="labelmess"></asp:Label> </td> </tr> </table> </ContentTemplate> </cc1:TabPanel> <cc1:TabPanel ID="TpMyPaymentsCk" runat="server" HeaderText="Payments Check" > <ContentTemplate> <asp:GridView ID="GdvmypaymentsCk" runat="server" CssClass="Mytabs" BackColor="White" BorderColor="White" BorderStyle="Ridge" BorderWidth="2px" CellPadding="3" CellSpacing="1" GridLines="None" OnPageIndexChanging="GdvmypaymentsCk_PageIndexChanging" OnSorting="GdvmypaymentsCk_Sorting" EmptyDataText="No Data"> <RowStyle BackColor="#DEDFDE" ForeColor="Black" /> <FooterStyle BackColor="#C6C3C6" ForeColor="Black" /> <PagerStyle BackColor="Control" ForeColor="Gray" HorizontalAlign="Left" /> <SelectedRowStyle BackColor="#9471DE" Font-Bold="True" ForeColor="White" /> <HeaderStyle BackColor="#4A3C8C" Font-Bold="True" ForeColor="#E7E7FF" /> <PagerSettings Position="TopAndBottom" /> </asp:GridView> </ContentTemplate> </cc1:TabPanel> <cc1:TabPanel ID="TpMyCalls" runat="server" HeaderText="Calls" > <ContentTemplate> <table> <tr> <td> <asp:GridView ID="GdvSelectcallsP" runat="server" CssClass="Mytabs" BackColor="White" BorderColor="White" BorderStyle="Ridge" BorderWidth="2px" CellPadding="3" CellSpacing="1" GridLines="None" OnPageIndexChanging="GdvSelectcallsP_PageIndexChanging" OnRowDataBound="GdvSelectcallsP_RowDataBound" OnSorting="GdvSelectcallsP_Sorting" > <RowStyle BackColor="#DEDFDE" ForeColor="Black" /> <FooterStyle BackColor="#C6C3C6" ForeColor="Black" /> <PagerStyle BackColor="#C6C3C6" ForeColor="Black" HorizontalAlign="Right" /> <SelectedRowStyle BackColor="#9471DE" Font-Bold="True" ForeColor="White" /> <HeaderStyle BackColor="#4A3C8C" Font-Bold="True" ForeColor="#E7E7FF" /> <Columns> <asp:TemplateField> <HeaderTemplate> Row No </HeaderTemplate> <ItemTemplate> <%# Container.DataItemIndex + 1 %> </ItemTemplate> </asp:TemplateField> </Columns> </asp:GridView> </td> </tr> <tr> <td> <asp:Label ID="lblmessselectcallsAhkP" runat="server" CssClass="labelmess"></asp:Label> </td> </tr> </table> </ContentTemplate> </cc1:TabPanel> </cc1:TabContainer> </asp:Content>

    Read the article

  • Cardinality Estimation Bug with Lookups in SQL Server 2008 onward

    - by Paul White
    Cost-based optimization stands or falls on the quality of cardinality estimates (expected row counts).  If the optimizer has incorrect information to start with, it is quite unlikely to produce good quality execution plans except by chance.  There are many ways we can provide good starting information to the optimizer, and even more ways for cardinality estimation to go wrong.  Good database people know this, and work hard to write optimizer-friendly queries with a schema and metadata (e.g. statistics) that reduce the chances of poor cardinality estimation producing a sub-optimal plan.  Today, I am going to look at a case where poor cardinality estimation is Microsoft’s fault, and not yours. SQL Server 2005 SELECT th.ProductID, th.TransactionID, th.TransactionDate FROM Production.TransactionHistory AS th WHERE th.ProductID = 1 AND th.TransactionDate BETWEEN '20030901' AND '20031231'; The query plan on SQL Server 2005 is as follows (if you are using a more recent version of AdventureWorks, you will need to change the year on the date range from 2003 to 2007): There is an Index Seek on ProductID = 1, followed by a Key Lookup to find the Transaction Date for each row, and finally a Filter to restrict the results to only those rows where Transaction Date falls in the range specified.  The cardinality estimate of 45 rows at the Index Seek is exactly correct.  The table is not very large, there are up-to-date statistics associated with the index, so this is as expected. The estimate for the Key Lookup is also exactly right.  Each lookup into the Clustered Index to find the Transaction Date is guaranteed to return exactly one row.  The plan shows that the Key Lookup is expected to be executed 45 times.  The estimate for the Inner Join output is also correct – 45 rows from the seek joining to one row each time, gives 45 rows as output. The Filter estimate is also very good: the optimizer estimates 16.9951 rows will match the specified range of transaction dates.  Eleven rows are produced by this query, but that small difference is quite normal and certainly nothing to worry about here.  All good so far. SQL Server 2008 onward The same query executed against an identical copy of AdventureWorks on SQL Server 2008 produces a different execution plan: The optimizer has pushed the Filter conditions seen in the 2005 plan down to the Key Lookup.  This is a good optimization – it makes sense to filter rows out as early as possible.  Unfortunately, it has made a bit of a mess of the cardinality estimates. The post-Filter estimate of 16.9951 rows seen in the 2005 plan has moved with the predicate on Transaction Date.  Instead of estimating one row, the plan now suggests that 16.9951 rows will be produced by each clustered index lookup – clearly not right!  This misinformation also confuses SQL Sentry Plan Explorer: Plan Explorer shows 765 rows expected from the Key Lookup (it multiplies a rounded estimate of 17 rows by 45 expected executions to give 765 rows total). Workarounds One workaround is to provide a covering non-clustered index (avoiding the lookup avoids the problem of course): CREATE INDEX nc1 ON Production.TransactionHistory (ProductID) INCLUDE (TransactionDate); With the Transaction Date filter applied as a residual predicate in the same operator as the seek, the estimate is again as expected: We could also force the use of the ultimate covering index (the clustered one): SELECT th.ProductID, th.TransactionID, th.TransactionDate FROM Production.TransactionHistory AS th WITH (INDEX(1)) WHERE th.ProductID = 1 AND th.TransactionDate BETWEEN '20030901' AND '20031231'; Summary Providing a covering non-clustered index for all possible queries is not always practical, and scanning the clustered index will rarely be optimal.  Nevertheless, these are the best workarounds we have today. In the meantime, watch out for poor cardinality estimates when a predicate is applied as part of a lookup. The worst thing is that the estimate after the lookup join in the 2008+ plans is wrong.  It’s not hopelessly wrong in this particular case (45 versus 16.9951 is not the end of the world) but it easily can be much worse, and there’s not much you can do about it.  Any decisions made by the optimizer after such a lookup could be based on very wrong information – which can only be bad news. If you think this situation should be improved, please vote for this Connect item. © 2012 Paul White – All Rights Reserved twitter: @SQL_Kiwi email: [email protected]

    Read the article

  • Deleting xml file using radio value

    - by ???? ???
    i using php to delete file, but i got table loop like this: <table border="0" width="100%" cellpadding="0" cellspacing="0" id="product-table"> <tr class="bg_tableheader"> <th class="table-header-check"><a id="toggle-all" ></a> </th> <th class="table-header-check"><a href="#"><font color="white">Username</font></a> </th> <th class="table-header-check"><a href="#"><font color="white">First Name</font></a></th> <th class="table-header-check"><a href="#"><font color="white">Last Name</font></a></th> <th class="table-header-check"><a href="#"><font color="white">Email</font></a></th> <th class="table-header-check"><a href="#"><font color="white">Group</font></a></th> <th class="table-header-check"><a href="#"><font color="white">Birthday</font></a></th> <th class="table-header-check"><a href="#"><font color="white">Gender</font></a></th> <th class="table-header-check"><a href="#"><font color="white">Age</font></a></th> <th class="table-header-check"><a href="#"><font color="white">Country</font></a></th> </tr> <?php $files = glob('users/*.xml'); foreach($files as $file){ $xml = new SimpleXMLElement($file, 0, true); echo ' <tr> <td></td> <form action="" method="post"> <td class="alternate-row1"><input type="radio" name="file_name" value="'. basename($file, '.xml') .'" />'. basename($file, '.xml') .'</td> <td>'. $xml->name .'&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp&nbsp</td> <td class="alternate-row1">'. $xml->lastname .'</td> <td>'. $xml->email .'</td> <td class="alternate-row1">'. $xml->level .'</td> <td>'. $xml->birthday .'</td> <td class="alternate-row1">'. $xml->gender .'</td> <td>'. $xml->age .'</td> <td class="alternate-row1">'. $xml->country .'</td> </tr>'; } ?> </table> </div> <?php if(isset($_POST['file_name'])){ unlink('users/'.$_POST['file_name']); } ?> <input type="submit" value="Delete" /> </form> so as you can see i got radio value set has basename (xml file name) but from some reason it not working, any idea why is that? Thanks in advance.

    Read the article

  • Fun with Aggregates

    - by Paul White
    There are interesting things to be learned from even the simplest queries.  For example, imagine you are given the task of writing a query to list AdventureWorks product names where the product has at least one entry in the transaction history table, but fewer than ten. One possible query to meet that specification is: SELECT p.Name FROM Production.Product AS p JOIN Production.TransactionHistory AS th ON p.ProductID = th.ProductID GROUP BY p.ProductID, p.Name HAVING COUNT_BIG(*) < 10; That query correctly returns 23 rows (execution plan and data sample shown below): The execution plan looks a bit different from the written form of the query: the base tables are accessed in reverse order, and the aggregation is performed before the join.  The general idea is to read all rows from the history table, compute the count of rows grouped by ProductID, merge join the results to the Product table on ProductID, and finally filter to only return rows where the count is less than ten. This ‘fully-optimized’ plan has an estimated cost of around 0.33 units.  The reason for the quote marks there is that this plan is not quite as optimal as it could be – surely it would make sense to push the Filter down past the join too?  To answer that, let’s look at some other ways to formulate this query.  This being SQL, there are any number of ways to write logically-equivalent query specifications, so we’ll just look at a couple of interesting ones.  The first query is an attempt to reverse-engineer T-SQL from the optimized query plan shown above.  It joins the result of pre-aggregating the history table to the Product table before filtering: SELECT p.Name FROM ( SELECT th.ProductID, cnt = COUNT_BIG(*) FROM Production.TransactionHistory AS th GROUP BY th.ProductID ) AS q1 JOIN Production.Product AS p ON p.ProductID = q1.ProductID WHERE q1.cnt < 10; Perhaps a little surprisingly, we get a slightly different execution plan: The results are the same (23 rows) but this time the Filter is pushed below the join!  The optimizer chooses nested loops for the join, because the cardinality estimate for rows passing the Filter is a bit low (estimate 1 versus 23 actual), though you can force a merge join with a hint and the Filter still appears below the join.  In yet another variation, the < 10 predicate can be ‘manually pushed’ by specifying it in a HAVING clause in the “q1” sub-query instead of in the WHERE clause as written above. The reason this predicate can be pushed past the join in this query form, but not in the original formulation is simply an optimizer limitation – it does make efforts (primarily during the simplification phase) to encourage logically-equivalent query specifications to produce the same execution plan, but the implementation is not completely comprehensive. Moving on to a second example, the following query specification results from phrasing the requirement as “list the products where there exists fewer than ten correlated rows in the history table”: SELECT p.Name FROM Production.Product AS p WHERE EXISTS ( SELECT * FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID HAVING COUNT_BIG(*) < 10 ); Unfortunately, this query produces an incorrect result (86 rows): The problem is that it lists products with no history rows, though the reasons are interesting.  The COUNT_BIG(*) in the EXISTS clause is a scalar aggregate (meaning there is no GROUP BY clause) and scalar aggregates always produce a value, even when the input is an empty set.  In the case of the COUNT aggregate, the result of aggregating the empty set is zero (the other standard aggregates produce a NULL).  To make the point really clear, let’s look at product 709, which happens to be one for which no history rows exist: -- Scalar aggregate SELECT COUNT_BIG(*) FROM Production.TransactionHistory AS th WHERE th.ProductID = 709;   -- Vector aggregate SELECT COUNT_BIG(*) FROM Production.TransactionHistory AS th WHERE th.ProductID = 709 GROUP BY th.ProductID; The estimated execution plans for these two statements are almost identical: You might expect the Stream Aggregate to have a Group By for the second statement, but this is not the case.  The query includes an equality comparison to a constant value (709), so all qualified rows are guaranteed to have the same value for ProductID and the Group By is optimized away. In fact there are some minor differences between the two plans (the first is auto-parameterized and qualifies for trivial plan, whereas the second is not auto-parameterized and requires cost-based optimization), but there is nothing to indicate that one is a scalar aggregate and the other is a vector aggregate.  This is something I would like to see exposed in show plan so I suggested it on Connect.  Anyway, the results of running the two queries show the difference at runtime: The scalar aggregate (no GROUP BY) returns a result of zero, whereas the vector aggregate (with a GROUP BY clause) returns nothing at all.  Returning to our EXISTS query, we could ‘fix’ it by changing the HAVING clause to reject rows where the scalar aggregate returns zero: SELECT p.Name FROM Production.Product AS p WHERE EXISTS ( SELECT * FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID HAVING COUNT_BIG(*) BETWEEN 1 AND 9 ); The query now returns the correct 23 rows: Unfortunately, the execution plan is less efficient now – it has an estimated cost of 0.78 compared to 0.33 for the earlier plans.  Let’s try adding a redundant GROUP BY instead of changing the HAVING clause: SELECT p.Name FROM Production.Product AS p WHERE EXISTS ( SELECT * FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID GROUP BY th.ProductID HAVING COUNT_BIG(*) < 10 ); Not only do we now get correct results (23 rows), this is the execution plan: I like to compare that plan to quantum physics: if you don’t find it shocking, you haven’t understood it properly :)  The simple addition of a redundant GROUP BY has resulted in the EXISTS form of the query being transformed into exactly the same optimal plan we found earlier.  What’s more, in SQL Server 2008 and later, we can replace the odd-looking GROUP BY with an explicit GROUP BY on the empty set: SELECT p.Name FROM Production.Product AS p WHERE EXISTS ( SELECT * FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID GROUP BY () HAVING COUNT_BIG(*) < 10 ); I offer that as an alternative because some people find it more intuitive (and it perhaps has more geek value too).  Whichever way you prefer, it’s rather satisfying to note that the result of the sub-query does not exist for a particular correlated value where a vector aggregate is used (the scalar COUNT aggregate always returns a value, even if zero, so it always ‘EXISTS’ regardless which ProductID is logically being evaluated). The following query forms also produce the optimal plan and correct results, so long as a vector aggregate is used (you can probably find more equivalent query forms): WHERE Clause SELECT p.Name FROM Production.Product AS p WHERE ( SELECT COUNT_BIG(*) FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID GROUP BY () ) < 10; APPLY SELECT p.Name FROM Production.Product AS p CROSS APPLY ( SELECT NULL FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID GROUP BY () HAVING COUNT_BIG(*) < 10 ) AS ca (dummy); FROM Clause SELECT q1.Name FROM ( SELECT p.Name, cnt = ( SELECT COUNT_BIG(*) FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID GROUP BY () ) FROM Production.Product AS p ) AS q1 WHERE q1.cnt < 10; This last example uses SUM(1) instead of COUNT and does not require a vector aggregate…you should be able to work out why :) SELECT q.Name FROM ( SELECT p.Name, cnt = ( SELECT SUM(1) FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID ) FROM Production.Product AS p ) AS q WHERE q.cnt < 10; The semantics of SQL aggregates are rather odd in places.  It definitely pays to get to know the rules, and to be careful to check whether your queries are using scalar or vector aggregates.  As we have seen, query plans do not show in which ‘mode’ an aggregate is running and getting it wrong can cause poor performance, wrong results, or both. © 2012 Paul White Twitter: @SQL_Kiwi email: [email protected]

    Read the article

  • I see no LOBs!

    - by Paul White
    Is it possible to see LOB (large object) logical reads from STATISTICS IO output on a table with no LOB columns? I was asked this question today by someone who had spent a good fraction of their afternoon trying to work out why this was occurring – even going so far as to re-run DBCC CHECKDB to see if any corruption had taken place.  The table in question wasn’t particularly pretty – it had grown somewhat organically over time, with new columns being added every so often as the need arose.  Nevertheless, it remained a simple structure with no LOB columns – no TEXT or IMAGE, no XML, no MAX types – nothing aside from ordinary INT, MONEY, VARCHAR, and DATETIME types.  To add to the air of mystery, not every query that ran against the table would report LOB logical reads – just sometimes – but when it did, the query often took much longer to execute. Ok, enough of the pre-amble.  I can’t reproduce the exact structure here, but the following script creates a table that will serve to demonstrate the effect: IF OBJECT_ID(N'dbo.Test', N'U') IS NOT NULL DROP TABLE dbo.Test GO CREATE TABLE dbo.Test ( row_id NUMERIC IDENTITY NOT NULL,   col01 NVARCHAR(450) NOT NULL, col02 NVARCHAR(450) NOT NULL, col03 NVARCHAR(450) NOT NULL, col04 NVARCHAR(450) NOT NULL, col05 NVARCHAR(450) NOT NULL, col06 NVARCHAR(450) NOT NULL, col07 NVARCHAR(450) NOT NULL, col08 NVARCHAR(450) NOT NULL, col09 NVARCHAR(450) NOT NULL, col10 NVARCHAR(450) NOT NULL, CONSTRAINT [PK dbo.Test row_id] PRIMARY KEY CLUSTERED (row_id) ) ; The next script loads the ten variable-length character columns with one-character strings in the first row, two-character strings in the second row, and so on down to the 450th row: WITH Numbers AS ( -- Generates numbers 1 - 450 inclusive SELECT TOP (450) n = ROW_NUMBER() OVER (ORDER BY (SELECT 0)) FROM master.sys.columns C1, master.sys.columns C2, master.sys.columns C3 ORDER BY n ASC ) INSERT dbo.Test WITH (TABLOCKX) SELECT REPLICATE(N'A', N.n), REPLICATE(N'B', N.n), REPLICATE(N'C', N.n), REPLICATE(N'D', N.n), REPLICATE(N'E', N.n), REPLICATE(N'F', N.n), REPLICATE(N'G', N.n), REPLICATE(N'H', N.n), REPLICATE(N'I', N.n), REPLICATE(N'J', N.n) FROM Numbers AS N ORDER BY N.n ASC ; Once those two scripts have run, the table contains 450 rows and 10 columns of data like this: Most of the time, when we query data from this table, we don’t see any LOB logical reads, for example: -- Find the maximum length of the data in -- column 5 for a range of rows SELECT result = MAX(DATALENGTH(T.col05)) FROM dbo.Test AS T WHERE row_id BETWEEN 50 AND 100 ; But with a different query… -- Read all the data in column 1 SELECT result = MAX(DATALENGTH(T.col01)) FROM dbo.Test AS T ; …suddenly we have 49 LOB logical reads, as well as the ‘normal’ logical reads we would expect. The Explanation If we had tried to create this table in SQL Server 2000, we would have received a warning message to say that future INSERT or UPDATE operations on the table might fail if the resulting row exceeded the in-row storage limit of 8060 bytes.  If we needed to store more data than would fit in an 8060 byte row (including internal overhead) we had to use a LOB column – TEXT, NTEXT, or IMAGE.  These special data types store the large data values in a separate structure, with just a small pointer left in the original row. Row Overflow SQL Server 2005 introduced a feature called row overflow, which allows one or more variable-length columns in a row to move to off-row storage if the data in a particular row would otherwise exceed 8060 bytes.  You no longer receive a warning when creating (or altering) a table that might need more than 8060 bytes of in-row storage; if SQL Server finds that it can no longer fit a variable-length column in a particular row, it will silently move one or more of these columns off the row into a separate allocation unit. Only variable-length columns can be moved in this way (for example the (N)VARCHAR, VARBINARY, and SQL_VARIANT types).  Fixed-length columns (like INTEGER and DATETIME for example) never move into ‘row overflow’ storage.  The decision to move a column off-row is done on a row-by-row basis – so data in a particular column might be stored in-row for some table records, and off-row for others. In general, if SQL Server finds that it needs to move a column into row-overflow storage, it moves the largest variable-length column record for that row.  Note that in the case of an UPDATE statement that results in the 8060 byte limit being exceeded, it might not be the column that grew that is moved! Sneaky LOBs Anyway, that’s all very interesting but I don’t want to get too carried away with the intricacies of row-overflow storage internals.  The point is that it is now possible to define a table with non-LOB columns that will silently exceed the old row-size limit and result in ordinary variable-length columns being moved to off-row storage.  Adding new columns to a table, expanding an existing column definition, or simply storing more data in a column than you used to – all these things can result in one or more variable-length columns being moved off the row. Note that row-overflow storage is logically quite different from old-style LOB and new-style MAX data type storage – individual variable-length columns are still limited to 8000 bytes each – you can just have more of them now.  Having said that, the physical mechanisms involved are very similar to full LOB storage – a column moved to row-overflow leaves a 24-byte pointer record in the row, and the ‘separate storage’ I have been talking about is structured very similarly to both old-style LOBs and new-style MAX types.  The disadvantages are also the same: when SQL Server needs a row-overflow column value it needs to follow the in-row pointer a navigate another chain of pages, just like retrieving a traditional LOB. And Finally… In the example script presented above, the rows with row_id values from 402 to 450 inclusive all exceed the total in-row storage limit of 8060 bytes.  A SELECT that references a column in one of those rows that has moved to off-row storage will incur one or more lob logical reads as the storage engine locates the data.  The results on your system might vary slightly depending on your settings, of course; but in my tests only column 1 in rows 402-450 moved off-row.  You might like to play around with the script – updating columns, changing data type lengths, and so on – to see the effect on lob logical reads and which columns get moved when.  You might even see row-overflow columns moving back in-row if they are updated to be smaller (hint: reduce the size of a column entry by at least 1000 bytes if you hope to see this). Be aware that SQL Server will not warn you when it moves ‘ordinary’ variable-length columns into overflow storage, and it can have dramatic effects on performance.  It makes more sense than ever to choose column data types sensibly.  If you make every column a VARCHAR(8000) or NVARCHAR(4000), and someone stores data that results in a row needing more than 8060 bytes, SQL Server might turn some of your column data into pseudo-LOBs – all without saying a word. Finally, some people make a distinction between ordinary LOBs (those that can hold up to 2GB of data) and the LOB-like structures created by row-overflow (where columns are still limited to 8000 bytes) by referring to row-overflow LOBs as SLOBs.  I find that quite appealing, but the ‘S’ stands for ‘small’, which makes expanding the whole acronym a little daft-sounding…small large objects anyone? © Paul White 2011 email: [email protected] twitter: @SQL_Kiwi

    Read the article

  • When is a Seek not a Seek?

    - by Paul White
    The following script creates a single-column clustered table containing the integers from 1 to 1,000 inclusive. IF OBJECT_ID(N'tempdb..#Test', N'U') IS NOT NULL DROP TABLE #Test ; GO CREATE TABLE #Test ( id INTEGER PRIMARY KEY CLUSTERED ); ; INSERT #Test (id) SELECT V.number FROM master.dbo.spt_values AS V WHERE V.[type] = N'P' AND V.number BETWEEN 1 AND 1000 ; Let’s say we need to find the rows with values from 100 to 170, excluding any values that divide exactly by 10.  One way to write that query would be: SELECT T.id FROM #Test AS T WHERE T.id IN ( 101,102,103,104,105,106,107,108,109, 111,112,113,114,115,116,117,118,119, 121,122,123,124,125,126,127,128,129, 131,132,133,134,135,136,137,138,139, 141,142,143,144,145,146,147,148,149, 151,152,153,154,155,156,157,158,159, 161,162,163,164,165,166,167,168,169 ) ; That query produces a pretty efficient-looking query plan: Knowing that the source column is defined as an INTEGER, we could also express the query this way: SELECT T.id FROM #Test AS T WHERE T.id >= 101 AND T.id <= 169 AND T.id % 10 > 0 ; We get a similar-looking plan: If you look closely, you might notice that the line connecting the two icons is a little thinner than before.  The first query is estimated to produce 61.9167 rows – very close to the 63 rows we know the query will return.  The second query presents a tougher challenge for SQL Server because it doesn’t know how to predict the selectivity of the modulo expression (T.id % 10 > 0).  Without that last line, the second query is estimated to produce 68.1667 rows – a slight overestimate.  Adding the opaque modulo expression results in SQL Server guessing at the selectivity.  As you may know, the selectivity guess for a greater-than operation is 30%, so the final estimate is 30% of 68.1667, which comes to 20.45 rows. The second difference is that the Clustered Index Seek is costed at 99% of the estimated total for the statement.  For some reason, the final SELECT operator is assigned a small cost of 0.0000484 units; I have absolutely no idea why this is so, or what it models.  Nevertheless, we can compare the total cost for both queries: the first one comes in at 0.0033501 units, and the second at 0.0034054.  The important point is that the second query is costed very slightly higher than the first, even though it is expected to produce many fewer rows (20.45 versus 61.9167). If you run the two queries, they produce exactly the same results, and both complete so quickly that it is impossible to measure CPU usage for a single execution.  We can, however, compare the I/O statistics for a single run by running the queries with STATISTICS IO ON: Table '#Test'. Scan count 63, logical reads 126, physical reads 0. Table '#Test'. Scan count 01, logical reads 002, physical reads 0. The query with the IN list uses 126 logical reads (and has a ‘scan count’ of 63), while the second query form completes with just 2 logical reads (and a ‘scan count’ of 1).  It is no coincidence that 126 = 63 * 2, by the way.  It is almost as if the first query is doing 63 seeks, compared to one for the second query. In fact, that is exactly what it is doing.  There is no indication of this in the graphical plan, or the tool-tip that appears when you hover your mouse over the Clustered Index Seek icon.  To see the 63 seek operations, you have click on the Seek icon and look in the Properties window (press F4, or right-click and choose from the menu): The Seek Predicates list shows a total of 63 seek operations – one for each of the values from the IN list contained in the first query.  I have expanded the first seek node to show the details; it is seeking down the clustered index to find the entry with the value 101.  Each of the other 62 nodes expands similarly, and the same information is contained (even more verbosely) in the XML form of the plan. Each of the 63 seek operations starts at the root of the clustered index B-tree and navigates down to the leaf page that contains the sought key value.  Our table is just large enough to need a separate root page, so each seek incurs 2 logical reads (one for the root, and one for the leaf).  We can see the index depth using the INDEXPROPERTY function, or by using the a DMV: SELECT S.index_type_desc, S.index_depth FROM sys.dm_db_index_physical_stats ( DB_ID(N'tempdb'), OBJECT_ID(N'tempdb..#Test', N'U'), 1, 1, DEFAULT ) AS S ; Let’s look now at the Properties window when the Clustered Index Seek from the second query is selected: There is just one seek operation, which starts at the root of the index and navigates the B-tree looking for the first key that matches the Start range condition (id >= 101).  It then continues to read records at the leaf level of the index (following links between leaf-level pages if necessary) until it finds a row that does not meet the End range condition (id <= 169).  Every row that meets the seek range condition is also tested against the Residual Predicate highlighted above (id % 10 > 0), and is only returned if it matches that as well. You will not be surprised that the single seek (with a range scan and residual predicate) is much more efficient than 63 singleton seeks.  It is not 63 times more efficient (as the logical reads comparison would suggest), but it is around three times faster.  Let’s run both query forms 10,000 times and measure the elapsed time: DECLARE @i INTEGER, @n INTEGER = 10000, @s DATETIME = GETDATE() ; SET NOCOUNT ON; SET STATISTICS XML OFF; ; WHILE @n > 0 BEGIN SELECT @i = T.id FROM #Test AS T WHERE T.id IN ( 101,102,103,104,105,106,107,108,109, 111,112,113,114,115,116,117,118,119, 121,122,123,124,125,126,127,128,129, 131,132,133,134,135,136,137,138,139, 141,142,143,144,145,146,147,148,149, 151,152,153,154,155,156,157,158,159, 161,162,163,164,165,166,167,168,169 ) ; SET @n -= 1; END ; PRINT DATEDIFF(MILLISECOND, @s, GETDATE()) ; GO DECLARE @i INTEGER, @n INTEGER = 10000, @s DATETIME = GETDATE() ; SET NOCOUNT ON ; WHILE @n > 0 BEGIN SELECT @i = T.id FROM #Test AS T WHERE T.id >= 101 AND T.id <= 169 AND T.id % 10 > 0 ; SET @n -= 1; END ; PRINT DATEDIFF(MILLISECOND, @s, GETDATE()) ; On my laptop, running SQL Server 2008 build 4272 (SP2 CU2), the IN form of the query takes around 830ms and the range query about 300ms.  The main point of this post is not performance, however – it is meant as an introduction to the next few parts in this mini-series that will continue to explore scans and seeks in detail. When is a seek not a seek?  When it is 63 seeks © Paul White 2011 email: [email protected] twitter: @SQL_kiwi

    Read the article

  • Fun with Aggregates

    - by Paul White
    There are interesting things to be learned from even the simplest queries.  For example, imagine you are given the task of writing a query to list AdventureWorks product names where the product has at least one entry in the transaction history table, but fewer than ten. One possible query to meet that specification is: SELECT p.Name FROM Production.Product AS p JOIN Production.TransactionHistory AS th ON p.ProductID = th.ProductID GROUP BY p.ProductID, p.Name HAVING COUNT_BIG(*) < 10; That query correctly returns 23 rows (execution plan and data sample shown below): The execution plan looks a bit different from the written form of the query: the base tables are accessed in reverse order, and the aggregation is performed before the join.  The general idea is to read all rows from the history table, compute the count of rows grouped by ProductID, merge join the results to the Product table on ProductID, and finally filter to only return rows where the count is less than ten. This ‘fully-optimized’ plan has an estimated cost of around 0.33 units.  The reason for the quote marks there is that this plan is not quite as optimal as it could be – surely it would make sense to push the Filter down past the join too?  To answer that, let’s look at some other ways to formulate this query.  This being SQL, there are any number of ways to write logically-equivalent query specifications, so we’ll just look at a couple of interesting ones.  The first query is an attempt to reverse-engineer T-SQL from the optimized query plan shown above.  It joins the result of pre-aggregating the history table to the Product table before filtering: SELECT p.Name FROM ( SELECT th.ProductID, cnt = COUNT_BIG(*) FROM Production.TransactionHistory AS th GROUP BY th.ProductID ) AS q1 JOIN Production.Product AS p ON p.ProductID = q1.ProductID WHERE q1.cnt < 10; Perhaps a little surprisingly, we get a slightly different execution plan: The results are the same (23 rows) but this time the Filter is pushed below the join!  The optimizer chooses nested loops for the join, because the cardinality estimate for rows passing the Filter is a bit low (estimate 1 versus 23 actual), though you can force a merge join with a hint and the Filter still appears below the join.  In yet another variation, the < 10 predicate can be ‘manually pushed’ by specifying it in a HAVING clause in the “q1” sub-query instead of in the WHERE clause as written above. The reason this predicate can be pushed past the join in this query form, but not in the original formulation is simply an optimizer limitation – it does make efforts (primarily during the simplification phase) to encourage logically-equivalent query specifications to produce the same execution plan, but the implementation is not completely comprehensive. Moving on to a second example, the following query specification results from phrasing the requirement as “list the products where there exists fewer than ten correlated rows in the history table”: SELECT p.Name FROM Production.Product AS p WHERE EXISTS ( SELECT * FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID HAVING COUNT_BIG(*) < 10 ); Unfortunately, this query produces an incorrect result (86 rows): The problem is that it lists products with no history rows, though the reasons are interesting.  The COUNT_BIG(*) in the EXISTS clause is a scalar aggregate (meaning there is no GROUP BY clause) and scalar aggregates always produce a value, even when the input is an empty set.  In the case of the COUNT aggregate, the result of aggregating the empty set is zero (the other standard aggregates produce a NULL).  To make the point really clear, let’s look at product 709, which happens to be one for which no history rows exist: -- Scalar aggregate SELECT COUNT_BIG(*) FROM Production.TransactionHistory AS th WHERE th.ProductID = 709;   -- Vector aggregate SELECT COUNT_BIG(*) FROM Production.TransactionHistory AS th WHERE th.ProductID = 709 GROUP BY th.ProductID; The estimated execution plans for these two statements are almost identical: You might expect the Stream Aggregate to have a Group By for the second statement, but this is not the case.  The query includes an equality comparison to a constant value (709), so all qualified rows are guaranteed to have the same value for ProductID and the Group By is optimized away. In fact there are some minor differences between the two plans (the first is auto-parameterized and qualifies for trivial plan, whereas the second is not auto-parameterized and requires cost-based optimization), but there is nothing to indicate that one is a scalar aggregate and the other is a vector aggregate.  This is something I would like to see exposed in show plan so I suggested it on Connect.  Anyway, the results of running the two queries show the difference at runtime: The scalar aggregate (no GROUP BY) returns a result of zero, whereas the vector aggregate (with a GROUP BY clause) returns nothing at all.  Returning to our EXISTS query, we could ‘fix’ it by changing the HAVING clause to reject rows where the scalar aggregate returns zero: SELECT p.Name FROM Production.Product AS p WHERE EXISTS ( SELECT * FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID HAVING COUNT_BIG(*) BETWEEN 1 AND 9 ); The query now returns the correct 23 rows: Unfortunately, the execution plan is less efficient now – it has an estimated cost of 0.78 compared to 0.33 for the earlier plans.  Let’s try adding a redundant GROUP BY instead of changing the HAVING clause: SELECT p.Name FROM Production.Product AS p WHERE EXISTS ( SELECT * FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID GROUP BY th.ProductID HAVING COUNT_BIG(*) < 10 ); Not only do we now get correct results (23 rows), this is the execution plan: I like to compare that plan to quantum physics: if you don’t find it shocking, you haven’t understood it properly :)  The simple addition of a redundant GROUP BY has resulted in the EXISTS form of the query being transformed into exactly the same optimal plan we found earlier.  What’s more, in SQL Server 2008 and later, we can replace the odd-looking GROUP BY with an explicit GROUP BY on the empty set: SELECT p.Name FROM Production.Product AS p WHERE EXISTS ( SELECT * FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID GROUP BY () HAVING COUNT_BIG(*) < 10 ); I offer that as an alternative because some people find it more intuitive (and it perhaps has more geek value too).  Whichever way you prefer, it’s rather satisfying to note that the result of the sub-query does not exist for a particular correlated value where a vector aggregate is used (the scalar COUNT aggregate always returns a value, even if zero, so it always ‘EXISTS’ regardless which ProductID is logically being evaluated). The following query forms also produce the optimal plan and correct results, so long as a vector aggregate is used (you can probably find more equivalent query forms): WHERE Clause SELECT p.Name FROM Production.Product AS p WHERE ( SELECT COUNT_BIG(*) FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID GROUP BY () ) < 10; APPLY SELECT p.Name FROM Production.Product AS p CROSS APPLY ( SELECT NULL FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID GROUP BY () HAVING COUNT_BIG(*) < 10 ) AS ca (dummy); FROM Clause SELECT q1.Name FROM ( SELECT p.Name, cnt = ( SELECT COUNT_BIG(*) FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID GROUP BY () ) FROM Production.Product AS p ) AS q1 WHERE q1.cnt < 10; This last example uses SUM(1) instead of COUNT and does not require a vector aggregate…you should be able to work out why :) SELECT q.Name FROM ( SELECT p.Name, cnt = ( SELECT SUM(1) FROM Production.TransactionHistory AS th WHERE th.ProductID = p.ProductID ) FROM Production.Product AS p ) AS q WHERE q.cnt < 10; The semantics of SQL aggregates are rather odd in places.  It definitely pays to get to know the rules, and to be careful to check whether your queries are using scalar or vector aggregates.  As we have seen, query plans do not show in which ‘mode’ an aggregate is running and getting it wrong can cause poor performance, wrong results, or both. © 2012 Paul White Twitter: @SQL_Kiwi email: [email protected]

    Read the article

  • Heaps of Trouble?

    - by Paul White NZ
    If you’re not already a regular reader of Brad Schulz’s blog, you’re missing out on some great material.  In his latest entry, he is tasked with optimizing a query run against tables that have no indexes at all.  The problem is, predictably, that performance is not very good.  The catch is that we are not allowed to create any indexes (or even new statistics) as part of our optimization efforts. In this post, I’m going to look at the problem from a slightly different angle, and present an alternative solution to the one Brad found.  Inevitably, there’s going to be some overlap between our entries, and while you don’t necessarily need to read Brad’s post before this one, I do strongly recommend that you read it at some stage; he covers some important points that I won’t cover again here. The Example We’ll use data from the AdventureWorks database, copied to temporary unindexed tables.  A script to create these structures is shown below: CREATE TABLE #Custs ( CustomerID INTEGER NOT NULL, TerritoryID INTEGER NULL, CustomerType NCHAR(1) COLLATE SQL_Latin1_General_CP1_CI_AI NOT NULL, ); GO CREATE TABLE #Prods ( ProductMainID INTEGER NOT NULL, ProductSubID INTEGER NOT NULL, ProductSubSubID INTEGER NOT NULL, Name NVARCHAR(50) COLLATE SQL_Latin1_General_CP1_CI_AI NOT NULL, ); GO CREATE TABLE #OrdHeader ( SalesOrderID INTEGER NOT NULL, OrderDate DATETIME NOT NULL, SalesOrderNumber NVARCHAR(25) COLLATE SQL_Latin1_General_CP1_CI_AI NOT NULL, CustomerID INTEGER NOT NULL, ); GO CREATE TABLE #OrdDetail ( SalesOrderID INTEGER NOT NULL, OrderQty SMALLINT NOT NULL, LineTotal NUMERIC(38,6) NOT NULL, ProductMainID INTEGER NOT NULL, ProductSubID INTEGER NOT NULL, ProductSubSubID INTEGER NOT NULL, ); GO INSERT #Custs ( CustomerID, TerritoryID, CustomerType ) SELECT C.CustomerID, C.TerritoryID, C.CustomerType FROM AdventureWorks.Sales.Customer C WITH (TABLOCK); GO INSERT #Prods ( ProductMainID, ProductSubID, ProductSubSubID, Name ) SELECT P.ProductID, P.ProductID, P.ProductID, P.Name FROM AdventureWorks.Production.Product P WITH (TABLOCK); GO INSERT #OrdHeader ( SalesOrderID, OrderDate, SalesOrderNumber, CustomerID ) SELECT H.SalesOrderID, H.OrderDate, H.SalesOrderNumber, H.CustomerID FROM AdventureWorks.Sales.SalesOrderHeader H WITH (TABLOCK); GO INSERT #OrdDetail ( SalesOrderID, OrderQty, LineTotal, ProductMainID, ProductSubID, ProductSubSubID ) SELECT D.SalesOrderID, D.OrderQty, D.LineTotal, D.ProductID, D.ProductID, D.ProductID FROM AdventureWorks.Sales.SalesOrderDetail D WITH (TABLOCK); The query itself is a simple join of the four tables: SELECT P.ProductMainID AS PID, P.Name, D.OrderQty, H.SalesOrderNumber, H.OrderDate, C.TerritoryID FROM #Prods P JOIN #OrdDetail D ON P.ProductMainID = D.ProductMainID AND P.ProductSubID = D.ProductSubID AND P.ProductSubSubID = D.ProductSubSubID JOIN #OrdHeader H ON D.SalesOrderID = H.SalesOrderID JOIN #Custs C ON H.CustomerID = C.CustomerID ORDER BY P.ProductMainID ASC OPTION (RECOMPILE, MAXDOP 1); Remember that these tables have no indexes at all, and only the single-column sampled statistics SQL Server automatically creates (assuming default settings).  The estimated query plan produced for the test query looks like this (click to enlarge): The Problem The problem here is one of cardinality estimation – the number of rows SQL Server expects to find at each step of the plan.  The lack of indexes and useful statistical information means that SQL Server does not have the information it needs to make a good estimate.  Every join in the plan shown above estimates that it will produce just a single row as output.  Brad covers the factors that lead to the low estimates in his post. In reality, the join between the #Prods and #OrdDetail tables will produce 121,317 rows.  It should not surprise you that this has rather dire consequences for the remainder of the query plan.  In particular, it makes a nonsense of the optimizer’s decision to use Nested Loops to join to the two remaining tables.  Instead of scanning the #OrdHeader and #Custs tables once (as it expected), it has to perform 121,317 full scans of each.  The query takes somewhere in the region of twenty minutes to run to completion on my development machine. A Solution At this point, you may be thinking the same thing I was: if we really are stuck with no indexes, the best we can do is to use hash joins everywhere. We can force the exclusive use of hash joins in several ways, the two most common being join and query hints.  A join hint means writing the query using the INNER HASH JOIN syntax; using a query hint involves adding OPTION (HASH JOIN) at the bottom of the query.  The difference is that using join hints also forces the order of the join, whereas the query hint gives the optimizer freedom to reorder the joins at its discretion. Adding the OPTION (HASH JOIN) hint results in this estimated plan: That produces the correct output in around seven seconds, which is quite an improvement!  As a purely practical matter, and given the rigid rules of the environment we find ourselves in, we might leave things there.  (We can improve the hashing solution a bit – I’ll come back to that later on). Faster Nested Loops It might surprise you to hear that we can beat the performance of the hash join solution shown above using nested loops joins exclusively, and without breaking the rules we have been set. The key to this part is to realize that a condition like (A = B) can be expressed as (A <= B) AND (A >= B).  Armed with this tremendous new insight, we can rewrite the join predicates like so: SELECT P.ProductMainID AS PID, P.Name, D.OrderQty, H.SalesOrderNumber, H.OrderDate, C.TerritoryID FROM #OrdDetail D JOIN #OrdHeader H ON D.SalesOrderID >= H.SalesOrderID AND D.SalesOrderID <= H.SalesOrderID JOIN #Custs C ON H.CustomerID >= C.CustomerID AND H.CustomerID <= C.CustomerID JOIN #Prods P ON P.ProductMainID >= D.ProductMainID AND P.ProductMainID <= D.ProductMainID AND P.ProductSubID = D.ProductSubID AND P.ProductSubSubID = D.ProductSubSubID ORDER BY D.ProductMainID OPTION (RECOMPILE, LOOP JOIN, MAXDOP 1, FORCE ORDER); I’ve also added LOOP JOIN and FORCE ORDER query hints to ensure that only nested loops joins are used, and that the tables are joined in the order they appear.  The new estimated execution plan is: This new query runs in under 2 seconds. Why Is It Faster? The main reason for the improvement is the appearance of the eager Index Spools, which are also known as index-on-the-fly spools.  If you read my Inside The Optimiser series you might be interested to know that the rule responsible is called JoinToIndexOnTheFly. An eager index spool consumes all rows from the table it sits above, and builds a index suitable for the join to seek on.  Taking the index spool above the #Custs table as an example, it reads all the CustomerID and TerritoryID values with a single scan of the table, and builds an index keyed on CustomerID.  The term ‘eager’ means that the spool consumes all of its input rows when it starts up.  The index is built in a work table in tempdb, has no associated statistics, and only exists until the query finishes executing. The result is that each unindexed table is only scanned once, and just for the columns necessary to build the temporary index.  From that point on, every execution of the inner side of the join is answered by a seek on the temporary index – not the base table. A second optimization is that the sort on ProductMainID (required by the ORDER BY clause) is performed early, on just the rows coming from the #OrdDetail table.  The optimizer has a good estimate for the number of rows it needs to sort at that stage – it is just the cardinality of the table itself.  The accuracy of the estimate there is important because it helps determine the memory grant given to the sort operation.  Nested loops join preserves the order of rows on its outer input, so sorting early is safe.  (Hash joins do not preserve order in this way, of course). The extra lazy spool on the #Prods branch is a further optimization that avoids executing the seek on the temporary index if the value being joined (the ‘outer reference’) hasn’t changed from the last row received on the outer input.  It takes advantage of the fact that rows are still sorted on ProductMainID, so if duplicates exist, they will arrive at the join operator one after the other. The optimizer is quite conservative about introducing index spools into a plan, because creating and dropping a temporary index is a relatively expensive operation.  It’s presence in a plan is often an indication that a useful index is missing. I want to stress that I rewrote the query in this way primarily as an educational exercise – I can’t imagine having to do something so horrible to a production system. Improving the Hash Join I promised I would return to the solution that uses hash joins.  You might be puzzled that SQL Server can create three new indexes (and perform all those nested loops iterations) faster than it can perform three hash joins.  The answer, again, is down to the poor information available to the optimizer.  Let’s look at the hash join plan again: Two of the hash joins have single-row estimates on their build inputs.  SQL Server fixes the amount of memory available for the hash table based on this cardinality estimate, so at run time the hash join very quickly runs out of memory. This results in the join spilling hash buckets to disk, and any rows from the probe input that hash to the spilled buckets also get written to disk.  The join process then continues, and may again run out of memory.  This is a recursive process, which may eventually result in SQL Server resorting to a bailout join algorithm, which is guaranteed to complete eventually, but may be very slow.  The data sizes in the example tables are not large enough to force a hash bailout, but it does result in multiple levels of hash recursion.  You can see this for yourself by tracing the Hash Warning event using the Profiler tool. The final sort in the plan also suffers from a similar problem: it receives very little memory and has to perform multiple sort passes, saving intermediate runs to disk (the Sort Warnings Profiler event can be used to confirm this).  Notice also that because hash joins don’t preserve sort order, the sort cannot be pushed down the plan toward the #OrdDetail table, as in the nested loops plan. Ok, so now we understand the problems, what can we do to fix it?  We can address the hash spilling by forcing a different order for the joins: SELECT P.ProductMainID AS PID, P.Name, D.OrderQty, H.SalesOrderNumber, H.OrderDate, C.TerritoryID FROM #Prods P JOIN #Custs C JOIN #OrdHeader H ON H.CustomerID = C.CustomerID JOIN #OrdDetail D ON D.SalesOrderID = H.SalesOrderID ON P.ProductMainID = D.ProductMainID AND P.ProductSubID = D.ProductSubID AND P.ProductSubSubID = D.ProductSubSubID ORDER BY D.ProductMainID OPTION (MAXDOP 1, HASH JOIN, FORCE ORDER); With this plan, each of the inputs to the hash joins has a good estimate, and no hash recursion occurs.  The final sort still suffers from the one-row estimate problem, and we get a single-pass sort warning as it writes rows to disk.  Even so, the query runs to completion in three or four seconds.  That’s around half the time of the previous hashing solution, but still not as fast as the nested loops trickery. Final Thoughts SQL Server’s optimizer makes cost-based decisions, so it is vital to provide it with accurate information.  We can’t really blame the performance problems highlighted here on anything other than the decision to use completely unindexed tables, and not to allow the creation of additional statistics. I should probably stress that the nested loops solution shown above is not one I would normally contemplate in the real world.  It’s there primarily for its educational and entertainment value.  I might perhaps use it to demonstrate to the sceptical that SQL Server itself is crying out for an index. Be sure to read Brad’s original post for more details.  My grateful thanks to him for granting permission to reuse some of his material. Paul White Email: [email protected] Twitter: @PaulWhiteNZ

    Read the article

  • So…is it a Seek or a Scan?

    - by Paul White
    You’re probably most familiar with the terms ‘Seek’ and ‘Scan’ from the graphical plans produced by SQL Server Management Studio (SSMS).  The image to the left shows the most common ones, with the three types of scan at the top, followed by four types of seek.  You might look to the SSMS tool-tip descriptions to explain the differences between them: Not hugely helpful are they?  Both mention scans and ranges (nothing about seeks) and the Index Seek description implies that it will not scan the index entirely (which isn’t necessarily true). Recall also yesterday’s post where we saw two Clustered Index Seek operations doing very different things.  The first Seek performed 63 single-row seeking operations; and the second performed a ‘Range Scan’ (more on those later in this post).  I hope you agree that those were two very different operations, and perhaps you are wondering why there aren’t different graphical plan icons for Range Scans and Seeks?  I have often wondered about that, and the first person to mention it after yesterday’s post was Erin Stellato (twitter | blog): Before we go on to make sense of all this, let’s look at another example of how SQL Server confusingly mixes the terms ‘Scan’ and ‘Seek’ in different contexts.  The diagram below shows a very simple heap table with two columns, one of which is the non-clustered Primary Key, and the other has a non-unique non-clustered index defined on it.  The right hand side of the diagram shows a simple query, it’s associated query plan, and a couple of extracts from the SSMS tool-tip and Properties windows. Notice the ‘scan direction’ entry in the Properties window snippet.  Is this a seek or a scan?  The different references to Scans and Seeks are even more pronounced in the XML plan output that the graphical plan is based on.  This fragment is what lies behind the single Index Seek icon shown above: You’ll find the same confusing references to Seeks and Scans throughout the product and its documentation. Making Sense of Seeks Let’s forget all about scans for a moment, and think purely about seeks.  Loosely speaking, a seek is the process of navigating an index B-tree to find a particular index record, most often at the leaf level.  A seek starts at the root and navigates down through the levels of the index to find the point of interest: Singleton Lookups The simplest sort of seek predicate performs this traversal to find (at most) a single record.  This is the case when we search for a single value using a unique index and an equality predicate.  It should be readily apparent that this type of search will either find one record, or none at all.  This operation is known as a singleton lookup.  Given the example table from before, the following query is an example of a singleton lookup seek: Sadly, there’s nothing in the graphical plan or XML output to show that this is a singleton lookup – you have to infer it from the fact that this is a single-value equality seek on a unique index.  The other common examples of a singleton lookup are bookmark lookups – both the RID and Key Lookup forms are singleton lookups (an RID lookup finds a single record in a heap from the unique row locator, and a Key Lookup does much the same thing on a clustered table).  If you happen to run your query with STATISTICS IO ON, you will notice that ‘Scan Count’ is always zero for a singleton lookup. Range Scans The other type of seek predicate is a ‘seek plus range scan’, which I will refer to simply as a range scan.  The seek operation makes an initial descent into the index structure to find the first leaf row that qualifies, and then performs a range scan (either backwards or forwards in the index) until it reaches the end of the scan range. The ability of a range scan to proceed in either direction comes about because index pages at the same level are connected by a doubly-linked list – each page has a pointer to the previous page (in logical key order) as well as a pointer to the following page.  The doubly-linked list is represented by the green and red dotted arrows in the index diagram presented earlier.  One subtle (but important) point is that the notion of a ‘forward’ or ‘backward’ scan applies to the logical key order defined when the index was built.  In the present case, the non-clustered primary key index was created as follows: CREATE TABLE dbo.Example ( key_col INTEGER NOT NULL, data INTEGER NOT NULL, CONSTRAINT [PK dbo.Example key_col] PRIMARY KEY NONCLUSTERED (key_col ASC) ) ; Notice that the primary key index specifies an ascending sort order for the single key column.  This means that a forward scan of the index will retrieve keys in ascending order, while a backward scan would retrieve keys in descending key order.  If the index had been created instead on key_col DESC, a forward scan would retrieve keys in descending order, and a backward scan would return keys in ascending order. A range scan seek predicate may have a Start condition, an End condition, or both.  Where one is missing, the scan starts (or ends) at one extreme end of the index, depending on the scan direction.  Some examples might help clarify that: the following diagram shows four queries, each of which performs a single seek against a column holding every integer from 1 to 100 inclusive.  The results from each query are shown in the blue columns, and relevant attributes from the Properties window appear on the right: Query 1 specifies that all key_col values less than 5 should be returned in ascending order.  The query plan achieves this by seeking to the start of the index leaf (there is no explicit starting value) and scanning forward until the End condition (key_col < 5) is no longer satisfied (SQL Server knows it can stop looking as soon as it finds a key_col value that isn’t less than 5 because all later index entries are guaranteed to sort higher). Query 2 asks for key_col values greater than 95, in descending order.  SQL Server returns these results by seeking to the end of the index, and scanning backwards (in descending key order) until it comes across a row that isn’t greater than 95.  Sharp-eyed readers may notice that the end-of-scan condition is shown as a Start range value.  This is a bug in the XML show plan which bubbles up to the Properties window – when a backward scan is performed, the roles of the Start and End values are reversed, but the plan does not reflect that.  Oh well. Query 3 looks for key_col values that are greater than or equal to 10, and less than 15, in ascending order.  This time, SQL Server seeks to the first index record that matches the Start condition (key_col >= 10) and then scans forward through the leaf pages until the End condition (key_col < 15) is no longer met. Query 4 performs much the same sort of operation as Query 3, but requests the output in descending order.  Again, we have to mentally reverse the Start and End conditions because of the bug, but otherwise the process is the same as always: SQL Server finds the highest-sorting record that meets the condition ‘key_col < 25’ and scans backward until ‘key_col >= 20’ is no longer true. One final point to note: seek operations always have the Ordered: True attribute.  This means that the operator always produces rows in a sorted order, either ascending or descending depending on how the index was defined, and whether the scan part of the operation is forward or backward.  You cannot rely on this sort order in your queries of course (you must always specify an ORDER BY clause if order is important) but SQL Server can make use of the sort order internally.  In the four queries above, the query optimizer was able to avoid an explicit Sort operator to honour the ORDER BY clause, for example. Multiple Seek Predicates As we saw yesterday, a single index seek plan operator can contain one or more seek predicates.  These seek predicates can either be all singleton seeks or all range scans – SQL Server does not mix them.  For example, you might expect the following query to contain two seek predicates, a singleton seek to find the single record in the unique index where key_col = 10, and a range scan to find the key_col values between 15 and 20: SELECT key_col FROM dbo.Example WHERE key_col = 10 OR key_col BETWEEN 15 AND 20 ORDER BY key_col ASC ; In fact, SQL Server transforms the singleton seek (key_col = 10) to the equivalent range scan, Start:[key_col >= 10], End:[key_col <= 10].  This allows both range scans to be evaluated by a single seek operator.  To be clear, this query results in two range scans: one from 10 to 10, and one from 15 to 20. Final Thoughts That’s it for today – tomorrow we’ll look at monitoring singleton lookups and range scans, and I’ll show you a seek on a heap table. Yes, a seek.  On a heap.  Not an index! If you would like to run the queries in this post for yourself, there’s a script below.  Thanks for reading! IF OBJECT_ID(N'dbo.Example', N'U') IS NOT NULL BEGIN DROP TABLE dbo.Example; END ; -- Test table is a heap -- Non-clustered primary key on 'key_col' CREATE TABLE dbo.Example ( key_col INTEGER NOT NULL, data INTEGER NOT NULL, CONSTRAINT [PK dbo.Example key_col] PRIMARY KEY NONCLUSTERED (key_col) ) ; -- Non-unique non-clustered index on the 'data' column CREATE NONCLUSTERED INDEX [IX dbo.Example data] ON dbo.Example (data) ; -- Add 100 rows INSERT dbo.Example WITH (TABLOCKX) ( key_col, data ) SELECT key_col = V.number, data = V.number FROM master.dbo.spt_values AS V WHERE V.[type] = N'P' AND V.number BETWEEN 1 AND 100 ; -- ================ -- Singleton lookup -- ================ ; -- Single value equality seek in a unique index -- Scan count = 0 when STATISTIS IO is ON -- Check the XML SHOWPLAN SELECT E.key_col FROM dbo.Example AS E WHERE E.key_col = 32 ; -- =========== -- Range Scans -- =========== ; -- Query 1 SELECT E.key_col FROM dbo.Example AS E WHERE E.key_col <= 5 ORDER BY E.key_col ASC ; -- Query 2 SELECT E.key_col FROM dbo.Example AS E WHERE E.key_col > 95 ORDER BY E.key_col DESC ; -- Query 3 SELECT E.key_col FROM dbo.Example AS E WHERE E.key_col >= 10 AND E.key_col < 15 ORDER BY E.key_col ASC ; -- Query 4 SELECT E.key_col FROM dbo.Example AS E WHERE E.key_col >= 20 AND E.key_col < 25 ORDER BY E.key_col DESC ; -- Final query (singleton + range = 2 range scans) SELECT E.key_col FROM dbo.Example AS E WHERE E.key_col = 10 OR E.key_col BETWEEN 15 AND 20 ORDER BY E.key_col ASC ; -- === TIDY UP === DROP TABLE dbo.Example; © 2011 Paul White email: [email protected] twitter: @SQL_Kiwi

    Read the article

  • Hello Operator, My Switch Is Bored

    - by Paul White
    This is a post for T-SQL Tuesday #43 hosted by my good friend Rob Farley. The topic this month is Plan Operators. I haven’t taken part in T-SQL Tuesday before, but I do like to write about execution plans, so this seemed like a good time to start. This post is in two parts. The first part is primarily an excuse to use a pretty bad play on words in the title of this blog post (if you’re too young to know what a telephone operator or a switchboard is, I hate you). The second part of the post looks at an invisible query plan operator (so to speak). 1. My Switch Is Bored Allow me to present the rare and interesting execution plan operator, Switch: Books Online has this to say about Switch: Following that description, I had a go at producing a Fast Forward Cursor plan that used the TOP operator, but had no luck. That may be due to my lack of skill with cursors, I’m not too sure. The only application of Switch in SQL Server 2012 that I am familiar with requires a local partitioned view: CREATE TABLE dbo.T1 (c1 int NOT NULL CHECK (c1 BETWEEN 00 AND 24)); CREATE TABLE dbo.T2 (c1 int NOT NULL CHECK (c1 BETWEEN 25 AND 49)); CREATE TABLE dbo.T3 (c1 int NOT NULL CHECK (c1 BETWEEN 50 AND 74)); CREATE TABLE dbo.T4 (c1 int NOT NULL CHECK (c1 BETWEEN 75 AND 99)); GO CREATE VIEW V1 AS SELECT c1 FROM dbo.T1 UNION ALL SELECT c1 FROM dbo.T2 UNION ALL SELECT c1 FROM dbo.T3 UNION ALL SELECT c1 FROM dbo.T4; Not only that, but it needs an updatable local partitioned view. We’ll need some primary keys to meet that requirement: ALTER TABLE dbo.T1 ADD CONSTRAINT PK_T1 PRIMARY KEY (c1);   ALTER TABLE dbo.T2 ADD CONSTRAINT PK_T2 PRIMARY KEY (c1);   ALTER TABLE dbo.T3 ADD CONSTRAINT PK_T3 PRIMARY KEY (c1);   ALTER TABLE dbo.T4 ADD CONSTRAINT PK_T4 PRIMARY KEY (c1); We also need an INSERT statement that references the view. Even more specifically, to see a Switch operator, we need to perform a single-row insert (multi-row inserts use a different plan shape): INSERT dbo.V1 (c1) VALUES (1); And now…the execution plan: The Constant Scan manufactures a single row with no columns. The Compute Scalar works out which partition of the view the new value should go in. The Assert checks that the computed partition number is not null (if it is, an error is returned). The Nested Loops Join executes exactly once, with the partition id as an outer reference (correlated parameter). The Switch operator checks the value of the parameter and executes the corresponding input only. If the partition id is 0, the uppermost Clustered Index Insert is executed, adding a row to table T1. If the partition id is 1, the next lower Clustered Index Insert is executed, adding a row to table T2…and so on. In case you were wondering, here’s a query and execution plan for a multi-row insert to the view: INSERT dbo.V1 (c1) VALUES (1), (2); Yuck! An Eager Table Spool and four Filters! I prefer the Switch plan. My guess is that almost all the old strategies that used a Switch operator have been replaced over time, using things like a regular Concatenation Union All combined with Start-Up Filters on its inputs. Other new (relative to the Switch operator) features like table partitioning have specific execution plan support that doesn’t need the Switch operator either. This feels like a bit of a shame, but perhaps it is just nostalgia on my part, it’s hard to know. Please do let me know if you encounter a query that can still use the Switch operator in 2012 – it must be very bored if this is the only possible modern usage! 2. Invisible Plan Operators The second part of this post uses an example based on a question Dave Ballantyne asked using the SQL Sentry Plan Explorer plan upload facility. If you haven’t tried that yet, make sure you’re on the latest version of the (free) Plan Explorer software, and then click the Post to SQLPerformance.com button. That will create a site question with the query plan attached (which can be anonymized if the plan contains sensitive information). Aaron Bertrand and I keep a close eye on questions there, so if you have ever wanted to ask a query plan question of either of us, that’s a good way to do it. The problem The issue I want to talk about revolves around a query issued against a calendar table. The script below creates a simplified version and adds 100 years of per-day information to it: USE tempdb; GO CREATE TABLE dbo.Calendar ( dt date NOT NULL, isWeekday bit NOT NULL, theYear smallint NOT NULL,   CONSTRAINT PK__dbo_Calendar_dt PRIMARY KEY CLUSTERED (dt) ); GO -- Monday is the first day of the week for me SET DATEFIRST 1;   -- Add 100 years of data INSERT dbo.Calendar WITH (TABLOCKX) (dt, isWeekday, theYear) SELECT CA.dt, isWeekday = CASE WHEN DATEPART(WEEKDAY, CA.dt) IN (6, 7) THEN 0 ELSE 1 END, theYear = YEAR(CA.dt) FROM Sandpit.dbo.Numbers AS N CROSS APPLY ( VALUES (DATEADD(DAY, N.n - 1, CONVERT(date, '01 Jan 2000', 113))) ) AS CA (dt) WHERE N.n BETWEEN 1 AND 36525; The following query counts the number of weekend days in 2013: SELECT Days = COUNT_BIG(*) FROM dbo.Calendar AS C WHERE theYear = 2013 AND isWeekday = 0; It returns the correct result (104) using the following execution plan: The query optimizer has managed to estimate the number of rows returned from the table exactly, based purely on the default statistics created separately on the two columns referenced in the query’s WHERE clause. (Well, almost exactly, the unrounded estimate is 104.289 rows.) There is already an invisible operator in this query plan – a Filter operator used to apply the WHERE clause predicates. We can see it by re-running the query with the enormously useful (but undocumented) trace flag 9130 enabled: Now we can see the full picture. The whole table is scanned, returning all 36,525 rows, before the Filter narrows that down to just the 104 we want. Without the trace flag, the Filter is incorporated in the Clustered Index Scan as a residual predicate. It is a little bit more efficient than using a separate operator, but residual predicates are still something you will want to avoid where possible. The estimates are still spot on though: Anyway, looking to improve the performance of this query, Dave added the following filtered index to the Calendar table: CREATE NONCLUSTERED INDEX Weekends ON dbo.Calendar(theYear) WHERE isWeekday = 0; The original query now produces a much more efficient plan: Unfortunately, the estimated number of rows produced by the seek is now wrong (365 instead of 104): What’s going on? The estimate was spot on before we added the index! Explanation You might want to grab a coffee for this bit. Using another trace flag or two (8606 and 8612) we can see that the cardinality estimates were exactly right initially: The highlighted information shows the initial cardinality estimates for the base table (36,525 rows), the result of applying the two relational selects in our WHERE clause (104 rows), and after performing the COUNT_BIG(*) group by aggregate (1 row). All of these are correct, but that was before cost-based optimization got involved :) Cost-based optimization When cost-based optimization starts up, the logical tree above is copied into a structure (the ‘memo’) that has one group per logical operation (roughly speaking). The logical read of the base table (LogOp_Get) ends up in group 7; the two predicates (LogOp_Select) end up in group 8 (with the details of the selections in subgroups 0-6). These two groups still have the correct cardinalities as trace flag 8608 output (initial memo contents) shows: During cost-based optimization, a rule called SelToIdxStrategy runs on group 8. It’s job is to match logical selections to indexable expressions (SARGs). It successfully matches the selections (theYear = 2013, is Weekday = 0) to the filtered index, and writes a new alternative into the memo structure. The new alternative is entered into group 8 as option 1 (option 0 was the original LogOp_Select): The new alternative is to do nothing (PhyOp_NOP = no operation), but to instead follow the new logical instructions listed below the NOP. The LogOp_GetIdx (full read of an index) goes into group 21, and the LogOp_SelectIdx (selection on an index) is placed in group 22, operating on the result of group 21. The definition of the comparison ‘the Year = 2013’ (ScaOp_Comp downwards) was already present in the memo starting at group 2, so no new memo groups are created for that. New Cardinality Estimates The new memo groups require two new cardinality estimates to be derived. First, LogOp_Idx (full read of the index) gets a predicted cardinality of 10,436. This number comes from the filtered index statistics: DBCC SHOW_STATISTICS (Calendar, Weekends) WITH STAT_HEADER; The second new cardinality derivation is for the LogOp_SelectIdx applying the predicate (theYear = 2013). To get a number for this, the cardinality estimator uses statistics for the column ‘theYear’, producing an estimate of 365 rows (there are 365 days in 2013!): DBCC SHOW_STATISTICS (Calendar, theYear) WITH HISTOGRAM; This is where the mistake happens. Cardinality estimation should have used the filtered index statistics here, to get an estimate of 104 rows: DBCC SHOW_STATISTICS (Calendar, Weekends) WITH HISTOGRAM; Unfortunately, the logic has lost sight of the link between the read of the filtered index (LogOp_GetIdx) in group 22, and the selection on that index (LogOp_SelectIdx) that it is deriving a cardinality estimate for, in group 21. The correct cardinality estimate (104 rows) is still present in the memo, attached to group 8, but that group now has a PhyOp_NOP implementation. Skipping over the rest of cost-based optimization (in a belated attempt at brevity) we can see the optimizer’s final output using trace flag 8607: This output shows the (incorrect, but understandable) 365 row estimate for the index range operation, and the correct 104 estimate still attached to its PhyOp_NOP. This tree still has to go through a few post-optimizer rewrites and ‘copy out’ from the memo structure into a tree suitable for the execution engine. One step in this process removes PhyOp_NOP, discarding its 104-row cardinality estimate as it does so. To finish this section on a more positive note, consider what happens if we add an OVER clause to the query aggregate. This isn’t intended to be a ‘fix’ of any sort, I just want to show you that the 104 estimate can survive and be used if later cardinality estimation needs it: SELECT Days = COUNT_BIG(*) OVER () FROM dbo.Calendar AS C WHERE theYear = 2013 AND isWeekday = 0; The estimated execution plan is: Note the 365 estimate at the Index Seek, but the 104 lives again at the Segment! We can imagine the lost predicate ‘isWeekday = 0’ as sitting between the seek and the segment in an invisible Filter operator that drops the estimate from 365 to 104. Even though the NOP group is removed after optimization (so we don’t see it in the execution plan) bear in mind that all cost-based choices were made with the 104-row memo group present, so although things look a bit odd, it shouldn’t affect the optimizer’s plan selection. I should also mention that we can work around the estimation issue by including the index’s filtering columns in the index key: CREATE NONCLUSTERED INDEX Weekends ON dbo.Calendar(theYear, isWeekday) WHERE isWeekday = 0 WITH (DROP_EXISTING = ON); There are some downsides to doing this, including that changes to the isWeekday column may now require Halloween Protection, but that is unlikely to be a big problem for a static calendar table ;)  With the updated index in place, the original query produces an execution plan with the correct cardinality estimation showing at the Index Seek: That’s all for today, remember to let me know about any Switch plans you come across on a modern instance of SQL Server! Finally, here are some other posts of mine that cover other plan operators: Segment and Sequence Project Common Subexpression Spools Why Plan Operators Run Backwards Row Goals and the Top Operator Hash Match Flow Distinct Top N Sort Index Spools and Page Splits Singleton and Range Seeks Bitmaps Hash Join Performance Compute Scalar © 2013 Paul White – All Rights Reserved Twitter: @SQL_Kiwi

    Read the article

  • Beware Sneaky Reads with Unique Indexes

    - by Paul White NZ
    A few days ago, Sandra Mueller (twitter | blog) asked a question using twitter’s #sqlhelp hash tag: “Might SQL Server retrieve (out-of-row) LOB data from a table, even if the column isn’t referenced in the query?” Leaving aside trivial cases (like selecting a computed column that does reference the LOB data), one might be tempted to say that no, SQL Server does not read data you haven’t asked for.  In general, that’s quite correct; however there are cases where SQL Server might sneakily retrieve a LOB column… Example Table Here’s a T-SQL script to create that table and populate it with 1,000 rows: CREATE TABLE dbo.LOBtest ( pk INTEGER IDENTITY NOT NULL, some_value INTEGER NULL, lob_data VARCHAR(MAX) NULL, another_column CHAR(5) NULL, CONSTRAINT [PK dbo.LOBtest pk] PRIMARY KEY CLUSTERED (pk ASC) ); GO DECLARE @Data VARCHAR(MAX); SET @Data = REPLICATE(CONVERT(VARCHAR(MAX), 'x'), 65540);   WITH Numbers (n) AS ( SELECT ROW_NUMBER() OVER (ORDER BY (SELECT 0)) FROM master.sys.columns C1, master.sys.columns C2 ) INSERT LOBtest WITH (TABLOCKX) ( some_value, lob_data ) SELECT TOP (1000) N.n, @Data FROM Numbers N WHERE N.n <= 1000; Test 1: A Simple Update Let’s run a query to subtract one from every value in the some_value column: UPDATE dbo.LOBtest WITH (TABLOCKX) SET some_value = some_value - 1; As you might expect, modifying this integer column in 1,000 rows doesn’t take very long, or use many resources.  The STATITICS IO and TIME output shows a total of 9 logical reads, and 25ms elapsed time.  The query plan is also very simple: Looking at the Clustered Index Scan, we can see that SQL Server only retrieves the pk and some_value columns during the scan: The pk column is needed by the Clustered Index Update operator to uniquely identify the row that is being changed.  The some_value column is used by the Compute Scalar to calculate the new value.  (In case you are wondering what the Top operator is for, it is used to enforce SET ROWCOUNT). Test 2: Simple Update with an Index Now let’s create a nonclustered index keyed on the some_value column, with lob_data as an included column: CREATE NONCLUSTERED INDEX [IX dbo.LOBtest some_value (lob_data)] ON dbo.LOBtest (some_value) INCLUDE ( lob_data ) WITH ( FILLFACTOR = 100, MAXDOP = 1, SORT_IN_TEMPDB = ON ); This is not a useful index for our simple update query; imagine that someone else created it for a different purpose.  Let’s run our update query again: UPDATE dbo.LOBtest WITH (TABLOCKX) SET some_value = some_value - 1; We find that it now requires 4,014 logical reads and the elapsed query time has increased to around 100ms.  The extra logical reads (4 per row) are an expected consequence of maintaining the nonclustered index. The query plan is very similar to before (click to enlarge): The Clustered Index Update operator picks up the extra work of maintaining the nonclustered index. The new Compute Scalar operators detect whether the value in the some_value column has actually been changed by the update.  SQL Server may be able to skip maintaining the nonclustered index if the value hasn’t changed (see my previous post on non-updating updates for details).  Our simple query does change the value of some_data in every row, so this optimization doesn’t add any value in this specific case. The output list of columns from the Clustered Index Scan hasn’t changed from the one shown previously: SQL Server still just reads the pk and some_data columns.  Cool. Overall then, adding the nonclustered index hasn’t had any startling effects, and the LOB column data still isn’t being read from the table.  Let’s see what happens if we make the nonclustered index unique. Test 3: Simple Update with a Unique Index Here’s the script to create a new unique index, and drop the old one: CREATE UNIQUE NONCLUSTERED INDEX [UQ dbo.LOBtest some_value (lob_data)] ON dbo.LOBtest (some_value) INCLUDE ( lob_data ) WITH ( FILLFACTOR = 100, MAXDOP = 1, SORT_IN_TEMPDB = ON ); GO DROP INDEX [IX dbo.LOBtest some_value (lob_data)] ON dbo.LOBtest; Remember that SQL Server only enforces uniqueness on index keys (the some_data column).  The lob_data column is simply stored at the leaf-level of the non-clustered index.  With that in mind, we might expect this change to make very little difference.  Let’s see: UPDATE dbo.LOBtest WITH (TABLOCKX) SET some_value = some_value - 1; Whoa!  Now look at the elapsed time and logical reads: Scan count 1, logical reads 2016, physical reads 0, read-ahead reads 0, lob logical reads 36015, lob physical reads 0, lob read-ahead reads 15992.   CPU time = 172 ms, elapsed time = 16172 ms. Even with all the data and index pages in memory, the query took over 16 seconds to update just 1,000 rows, performing over 52,000 LOB logical reads (nearly 16,000 of those using read-ahead). Why on earth is SQL Server reading LOB data in a query that only updates a single integer column? The Query Plan The query plan for test 3 looks a bit more complex than before: In fact, the bottom level is exactly the same as we saw with the non-unique index.  The top level has heaps of new stuff though, which I’ll come to in a moment. You might be expecting to find that the Clustered Index Scan is now reading the lob_data column (for some reason).  After all, we need to explain where all the LOB logical reads are coming from.  Sadly, when we look at the properties of the Clustered Index Scan, we see exactly the same as before: SQL Server is still only reading the pk and some_value columns – so what’s doing the LOB reads? Updates that Sneakily Read Data We have to go as far as the Clustered Index Update operator before we see LOB data in the output list: [Expr1020] is a bit flag added by an earlier Compute Scalar.  It is set true if the some_value column has not been changed (part of the non-updating updates optimization I mentioned earlier). The Clustered Index Update operator adds two new columns: the lob_data column, and some_value_OLD.  The some_value_OLD column, as the name suggests, is the pre-update value of the some_value column.  At this point, the clustered index has already been updated with the new value, but we haven’t touched the nonclustered index yet. An interesting observation here is that the Clustered Index Update operator can read a column into the data flow as part of its update operation.  SQL Server could have read the LOB data as part of the initial Clustered Index Scan, but that would mean carrying the data through all the operations that occur prior to the Clustered Index Update.  The server knows it will have to go back to the clustered index row to update it, so it delays reading the LOB data until then.  Sneaky! Why the LOB Data Is Needed This is all very interesting (I hope), but why is SQL Server reading the LOB data?  For that matter, why does it need to pass the pre-update value of the some_value column out of the Clustered Index Update? The answer relates to the top row of the query plan for test 3.  I’ll reproduce it here for convenience: Notice that this is a wide (per-index) update plan.  SQL Server used a narrow (per-row) update plan in test 2, where the Clustered Index Update took care of maintaining the nonclustered index too.  I’ll talk more about this difference shortly. The Split/Sort/Collapse combination is an optimization, which aims to make per-index update plans more efficient.  It does this by breaking each update into a delete/insert pair, reordering the operations, removing any redundant operations, and finally applying the net effect of all the changes to the nonclustered index. Imagine we had a unique index which currently holds three rows with the values 1, 2, and 3.  If we run a query that adds 1 to each row value, we would end up with values 2, 3, and 4.  The net effect of all the changes is the same as if we simply deleted the value 1, and added a new value 4. By applying net changes, SQL Server can also avoid false unique-key violations.  If we tried to immediately update the value 1 to a 2, it would conflict with the existing value 2 (which would soon be updated to 3 of course) and the query would fail.  You might argue that SQL Server could avoid the uniqueness violation by starting with the highest value (3) and working down.  That’s fine, but it’s not possible to generalize this logic to work with every possible update query. SQL Server has to use a wide update plan if it sees any risk of false uniqueness violations.  It’s worth noting that the logic SQL Server uses to detect whether these violations are possible has definite limits.  As a result, you will often receive a wide update plan, even when you can see that no violations are possible. Another benefit of this optimization is that it includes a sort on the index key as part of its work.  Processing the index changes in index key order promotes sequential I/O against the nonclustered index. A side-effect of all this is that the net changes might include one or more inserts.  In order to insert a new row in the index, SQL Server obviously needs all the columns – the key column and the included LOB column.  This is the reason SQL Server reads the LOB data as part of the Clustered Index Update. In addition, the some_value_OLD column is required by the Split operator (it turns updates into delete/insert pairs).  In order to generate the correct index key delete operation, it needs the old key value. The irony is that in this case the Split/Sort/Collapse optimization is anything but.  Reading all that LOB data is extremely expensive, so it is sad that the current version of SQL Server has no way to avoid it. Finally, for completeness, I should mention that the Filter operator is there to filter out the non-updating updates. Beating the Set-Based Update with a Cursor One situation where SQL Server can see that false unique-key violations aren’t possible is where it can guarantee that only one row is being updated.  Armed with this knowledge, we can write a cursor (or the WHILE-loop equivalent) that updates one row at a time, and so avoids reading the LOB data: SET NOCOUNT ON; SET STATISTICS XML, IO, TIME OFF;   DECLARE @PK INTEGER, @StartTime DATETIME; SET @StartTime = GETUTCDATE();   DECLARE curUpdate CURSOR LOCAL FORWARD_ONLY KEYSET SCROLL_LOCKS FOR SELECT L.pk FROM LOBtest L ORDER BY L.pk ASC;   OPEN curUpdate;   WHILE (1 = 1) BEGIN FETCH NEXT FROM curUpdate INTO @PK;   IF @@FETCH_STATUS = -1 BREAK; IF @@FETCH_STATUS = -2 CONTINUE;   UPDATE dbo.LOBtest SET some_value = some_value - 1 WHERE CURRENT OF curUpdate; END;   CLOSE curUpdate; DEALLOCATE curUpdate;   SELECT DATEDIFF(MILLISECOND, @StartTime, GETUTCDATE()); That completes the update in 1280 milliseconds (remember test 3 took over 16 seconds!) I used the WHERE CURRENT OF syntax there and a KEYSET cursor, just for the fun of it.  One could just as well use a WHERE clause that specified the primary key value instead. Clustered Indexes A clustered index is the ultimate index with included columns: all non-key columns are included columns in a clustered index.  Let’s re-create the test table and data with an updatable primary key, and without any non-clustered indexes: IF OBJECT_ID(N'dbo.LOBtest', N'U') IS NOT NULL DROP TABLE dbo.LOBtest; GO CREATE TABLE dbo.LOBtest ( pk INTEGER NOT NULL, some_value INTEGER NULL, lob_data VARCHAR(MAX) NULL, another_column CHAR(5) NULL, CONSTRAINT [PK dbo.LOBtest pk] PRIMARY KEY CLUSTERED (pk ASC) ); GO DECLARE @Data VARCHAR(MAX); SET @Data = REPLICATE(CONVERT(VARCHAR(MAX), 'x'), 65540);   WITH Numbers (n) AS ( SELECT ROW_NUMBER() OVER (ORDER BY (SELECT 0)) FROM master.sys.columns C1, master.sys.columns C2 ) INSERT LOBtest WITH (TABLOCKX) ( pk, some_value, lob_data ) SELECT TOP (1000) N.n, N.n, @Data FROM Numbers N WHERE N.n <= 1000; Now here’s a query to modify the cluster keys: UPDATE dbo.LOBtest SET pk = pk + 1; The query plan is: As you can see, the Split/Sort/Collapse optimization is present, and we also gain an Eager Table Spool, for Halloween protection.  In addition, SQL Server now has no choice but to read the LOB data in the Clustered Index Scan: The performance is not great, as you might expect (even though there is no non-clustered index to maintain): Table 'LOBtest'. Scan count 1, logical reads 2011, physical reads 0, read-ahead reads 0, lob logical reads 36015, lob physical reads 0, lob read-ahead reads 15992.   Table 'Worktable'. Scan count 1, logical reads 2040, physical reads 0, read-ahead reads 0, lob logical reads 34000, lob physical reads 0, lob read-ahead reads 8000.   SQL Server Execution Times: CPU time = 483 ms, elapsed time = 17884 ms. Notice how the LOB data is read twice: once from the Clustered Index Scan, and again from the work table in tempdb used by the Eager Spool. If you try the same test with a non-unique clustered index (rather than a primary key), you’ll get a much more efficient plan that just passes the cluster key (including uniqueifier) around (no LOB data or other non-key columns): A unique non-clustered index (on a heap) works well too: Both those queries complete in a few tens of milliseconds, with no LOB reads, and just a few thousand logical reads.  (In fact the heap is rather more efficient). There are lots more fun combinations to try that I don’t have space for here. Final Thoughts The behaviour shown in this post is not limited to LOB data by any means.  If the conditions are met, any unique index that has included columns can produce similar behaviour – something to bear in mind when adding large INCLUDE columns to achieve covering queries, perhaps. Paul White Email: [email protected] Twitter: @PaulWhiteNZ

    Read the article

  • MERGE Bug with Filtered Indexes

    - by Paul White
    A MERGE statement can fail, and incorrectly report a unique key violation when: The target table uses a unique filtered index; and No key column of the filtered index is updated; and A column from the filtering condition is updated; and Transient key violations are possible Example Tables Say we have two tables, one that is the target of a MERGE statement, and another that contains updates to be applied to the target.  The target table contains three columns, an integer primary key, a single character alternate key, and a status code column.  A filtered unique index exists on the alternate key, but is only enforced where the status code is ‘a’: CREATE TABLE #Target ( pk integer NOT NULL, ak character(1) NOT NULL, status_code character(1) NOT NULL,   PRIMARY KEY (pk) );   CREATE UNIQUE INDEX uq1 ON #Target (ak) INCLUDE (status_code) WHERE status_code = 'a'; The changes table contains just an integer primary key (to identify the target row to change) and the new status code: CREATE TABLE #Changes ( pk integer NOT NULL, status_code character(1) NOT NULL,   PRIMARY KEY (pk) ); Sample Data The sample data for the example is: INSERT #Target (pk, ak, status_code) VALUES (1, 'A', 'a'), (2, 'B', 'a'), (3, 'C', 'a'), (4, 'A', 'd');   INSERT #Changes (pk, status_code) VALUES (1, 'd'), (4, 'a');          Target                     Changes +-----------------------+    +------------------+ ¦ pk ¦ ak ¦ status_code ¦    ¦ pk ¦ status_code ¦ ¦----+----+-------------¦    ¦----+-------------¦ ¦  1 ¦ A  ¦ a           ¦    ¦  1 ¦ d           ¦ ¦  2 ¦ B  ¦ a           ¦    ¦  4 ¦ a           ¦ ¦  3 ¦ C  ¦ a           ¦    +------------------+ ¦  4 ¦ A  ¦ d           ¦ +-----------------------+ The target table’s alternate key (ak) column is unique, for rows where status_code = ‘a’.  Applying the changes to the target will change row 1 from status ‘a’ to status ‘d’, and row 4 from status ‘d’ to status ‘a’.  The result of applying all the changes will still satisfy the filtered unique index, because the ‘A’ in row 1 will be deleted from the index and the ‘A’ in row 4 will be added. Merge Test One Let’s now execute a MERGE statement to apply the changes: MERGE #Target AS t USING #Changes AS c ON c.pk = t.pk WHEN MATCHED AND c.status_code <> t.status_code THEN UPDATE SET status_code = c.status_code; The MERGE changes the two target rows as expected.  The updated target table now contains: +-----------------------+ ¦ pk ¦ ak ¦ status_code ¦ ¦----+----+-------------¦ ¦  1 ¦ A  ¦ d           ¦ <—changed from ‘a’ ¦  2 ¦ B  ¦ a           ¦ ¦  3 ¦ C  ¦ a           ¦ ¦  4 ¦ A  ¦ a           ¦ <—changed from ‘d’ +-----------------------+ Merge Test Two Now let’s repopulate the changes table to reverse the updates we just performed: TRUNCATE TABLE #Changes;   INSERT #Changes (pk, status_code) VALUES (1, 'a'), (4, 'd'); This will change row 1 back to status ‘a’ and row 4 back to status ‘d’.  As a reminder, the current state of the tables is:          Target                        Changes +-----------------------+    +------------------+ ¦ pk ¦ ak ¦ status_code ¦    ¦ pk ¦ status_code ¦ ¦----+----+-------------¦    ¦----+-------------¦ ¦  1 ¦ A  ¦ d           ¦    ¦  1 ¦ a           ¦ ¦  2 ¦ B  ¦ a           ¦    ¦  4 ¦ d           ¦ ¦  3 ¦ C  ¦ a           ¦    +------------------+ ¦  4 ¦ A  ¦ a           ¦ +-----------------------+ We execute the same MERGE statement: MERGE #Target AS t USING #Changes AS c ON c.pk = t.pk WHEN MATCHED AND c.status_code <> t.status_code THEN UPDATE SET status_code = c.status_code; However this time we receive the following message: Msg 2601, Level 14, State 1, Line 1 Cannot insert duplicate key row in object 'dbo.#Target' with unique index 'uq1'. The duplicate key value is (A). The statement has been terminated. Applying the changes using UPDATE Let’s now rewrite the MERGE to use UPDATE instead: UPDATE t SET status_code = c.status_code FROM #Target AS t JOIN #Changes AS c ON t.pk = c.pk WHERE c.status_code <> t.status_code; This query succeeds where the MERGE failed.  The two rows are updated as expected: +-----------------------+ ¦ pk ¦ ak ¦ status_code ¦ ¦----+----+-------------¦ ¦  1 ¦ A  ¦ a           ¦ <—changed back to ‘a’ ¦  2 ¦ B  ¦ a           ¦ ¦  3 ¦ C  ¦ a           ¦ ¦  4 ¦ A  ¦ d           ¦ <—changed back to ‘d’ +-----------------------+ What went wrong with the MERGE? In this test, the MERGE query execution happens to apply the changes in the order of the ‘pk’ column. In test one, this was not a problem: row 1 is removed from the unique filtered index by changing status_code from ‘a’ to ‘d’ before row 4 is added.  At no point does the table contain two rows where ak = ‘A’ and status_code = ‘a’. In test two, however, the first change was to change row 1 from status ‘d’ to status ‘a’.  This change means there would be two rows in the filtered unique index where ak = ‘A’ (both row 1 and row 4 meet the index filtering criteria ‘status_code = a’). The storage engine does not allow the query processor to violate a unique key (unless IGNORE_DUP_KEY is ON, but that is a different story, and doesn’t apply to MERGE in any case).  This strict rule applies regardless of the fact that if all changes were applied, there would be no unique key violation (row 4 would eventually be changed from ‘a’ to ‘d’, removing it from the filtered unique index, and resolving the key violation). Why it went wrong The query optimizer usually detects when this sort of temporary uniqueness violation could occur, and builds a plan that avoids the issue.  I wrote about this a couple of years ago in my post Beware Sneaky Reads with Unique Indexes (you can read more about the details on pages 495-497 of Microsoft SQL Server 2008 Internals or in Craig Freedman’s blog post on maintaining unique indexes).  To summarize though, the optimizer introduces Split, Filter, Sort, and Collapse operators into the query plan to: Split each row update into delete followed by an inserts Filter out rows that would not change the index (due to the filter on the index, or a non-updating update) Sort the resulting stream by index key, with deletes before inserts Collapse delete/insert pairs on the same index key back into an update The effect of all this is that only net changes are applied to an index (as one or more insert, update, and/or delete operations).  In this case, the net effect is a single update of the filtered unique index: changing the row for ak = ‘A’ from pk = 4 to pk = 1.  In case that is less than 100% clear, let’s look at the operation in test two again:          Target                     Changes                   Result +-----------------------+    +------------------+    +-----------------------+ ¦ pk ¦ ak ¦ status_code ¦    ¦ pk ¦ status_code ¦    ¦ pk ¦ ak ¦ status_code ¦ ¦----+----+-------------¦    ¦----+-------------¦    ¦----+----+-------------¦ ¦  1 ¦ A  ¦ d           ¦    ¦  1 ¦ d           ¦    ¦  1 ¦ A  ¦ a           ¦ ¦  2 ¦ B  ¦ a           ¦    ¦  4 ¦ a           ¦    ¦  2 ¦ B  ¦ a           ¦ ¦  3 ¦ C  ¦ a           ¦    +------------------+    ¦  3 ¦ C  ¦ a           ¦ ¦  4 ¦ A  ¦ a           ¦                            ¦  4 ¦ A  ¦ d           ¦ +-----------------------+                            +-----------------------+ From the filtered index’s point of view (filtered for status_code = ‘a’ and shown in nonclustered index key order) the overall effect of the query is:   Before           After +---------+    +---------+ ¦ pk ¦ ak ¦    ¦ pk ¦ ak ¦ ¦----+----¦    ¦----+----¦ ¦  4 ¦ A  ¦    ¦  1 ¦ A  ¦ ¦  2 ¦ B  ¦    ¦  2 ¦ B  ¦ ¦  3 ¦ C  ¦    ¦  3 ¦ C  ¦ +---------+    +---------+ The single net change there is a change of pk from 4 to 1 for the nonclustered index entry ak = ‘A’.  This is the magic performed by the split, sort, and collapse.  Notice in particular how the original changes to the index key (on the ‘ak’ column) have been transformed into an update of a non-key column (pk is included in the nonclustered index).  By not updating any nonclustered index keys, we are guaranteed to avoid transient key violations. The Execution Plans The estimated MERGE execution plan that produces the incorrect key-violation error looks like this (click to enlarge in a new window): The successful UPDATE execution plan is (click to enlarge in a new window): The MERGE execution plan is a narrow (per-row) update.  The single Clustered Index Merge operator maintains both the clustered index and the filtered nonclustered index.  The UPDATE plan is a wide (per-index) update.  The clustered index is maintained first, then the Split, Filter, Sort, Collapse sequence is applied before the nonclustered index is separately maintained. There is always a wide update plan for any query that modifies the database. The narrow form is a performance optimization where the number of rows is expected to be relatively small, and is not available for all operations.  One of the operations that should disallow a narrow plan is maintaining a unique index where intermediate key violations could occur. Workarounds The MERGE can be made to work (producing a wide update plan with split, sort, and collapse) by: Adding all columns referenced in the filtered index’s WHERE clause to the index key (INCLUDE is not sufficient); or Executing the query with trace flag 8790 set e.g. OPTION (QUERYTRACEON 8790). Undocumented trace flag 8790 forces a wide update plan for any data-changing query (remember that a wide update plan is always possible).  Either change will produce a successfully-executing wide update plan for the MERGE that failed previously. Conclusion The optimizer fails to spot the possibility of transient unique key violations with MERGE under the conditions listed at the start of this post.  It incorrectly chooses a narrow plan for the MERGE, which cannot provide the protection of a split/sort/collapse sequence for the nonclustered index maintenance. The MERGE plan may fail at execution time depending on the order in which rows are processed, and the distribution of data in the database.  Worse, a previously solid MERGE query may suddenly start to fail unpredictably if a filtered unique index is added to the merge target table at any point. Connect bug filed here Tests performed on SQL Server 2012 SP1 CUI (build 11.0.3321) x64 Developer Edition © 2012 Paul White – All Rights Reserved Twitter: @SQL_Kiwi Email: [email protected]

    Read the article

  • Seeking on a Heap, and Two Useful DMVs

    - by Paul White
    So far in this mini-series on seeks and scans, we have seen that a simple ‘seek’ operation can be much more complex than it first appears.  A seek can contain one or more seek predicates – each of which can either identify at most one row in a unique index (a singleton lookup) or a range of values (a range scan).  When looking at a query plan, we will often need to look at the details of the seek operator in the Properties window to see how many operations it is performing, and what type of operation each one is.  As you saw in the first post in this series, the number of hidden seeking operations can have an appreciable impact on performance. Measuring Seeks and Scans I mentioned in my last post that there is no way to tell from a graphical query plan whether you are seeing a singleton lookup or a range scan.  You can work it out – if you happen to know that the index is defined as unique and the seek predicate is an equality comparison, but there’s no separate property that says ‘singleton lookup’ or ‘range scan’.  This is a shame, and if I had my way, the query plan would show different icons for range scans and singleton lookups – perhaps also indicating whether the operation was one or more of those operations underneath the covers. In light of all that, you might be wondering if there is another way to measure how many seeks of either type are occurring in your system, or for a particular query.  As is often the case, the answer is yes – we can use a couple of dynamic management views (DMVs): sys.dm_db_index_usage_stats and sys.dm_db_index_operational_stats. Index Usage Stats The index usage stats DMV contains counts of index operations from the perspective of the Query Executor (QE) – the SQL Server component that is responsible for executing the query plan.  It has three columns that are of particular interest to us: user_seeks – the number of times an Index Seek operator appears in an executed plan user_scans – the number of times a Table Scan or Index Scan operator appears in an executed plan user_lookups – the number of times an RID or Key Lookup operator appears in an executed plan An operator is counted once per execution (generating an estimated plan does not affect the totals), so an Index Seek that executes 10,000 times in a single plan execution adds 1 to the count of user seeks.  Even less intuitively, an operator is also counted once per execution even if it is not executed at all.  I will show you a demonstration of each of these things later in this post. Index Operational Stats The index operational stats DMV contains counts of index and table operations from the perspective of the Storage Engine (SE).  It contains a wealth of interesting information, but the two columns of interest to us right now are: range_scan_count – the number of range scans (including unrestricted full scans) on a heap or index structure singleton_lookup_count – the number of singleton lookups in a heap or index structure This DMV counts each SE operation, so 10,000 singleton lookups will add 10,000 to the singleton lookup count column, and a table scan that is executed 5 times will add 5 to the range scan count. The Test Rig To explore the behaviour of seeks and scans in detail, we will need to create a test environment.  The scripts presented here are best run on SQL Server 2008 Developer Edition, but the majority of the tests will work just fine on SQL Server 2005.  A couple of tests use partitioning, but these will be skipped if you are not running an Enterprise-equivalent SKU.  Ok, first up we need a database: USE master; GO IF DB_ID('ScansAndSeeks') IS NOT NULL DROP DATABASE ScansAndSeeks; GO CREATE DATABASE ScansAndSeeks; GO USE ScansAndSeeks; GO ALTER DATABASE ScansAndSeeks SET ALLOW_SNAPSHOT_ISOLATION OFF ; ALTER DATABASE ScansAndSeeks SET AUTO_CLOSE OFF, AUTO_SHRINK OFF, AUTO_CREATE_STATISTICS OFF, AUTO_UPDATE_STATISTICS OFF, PARAMETERIZATION SIMPLE, READ_COMMITTED_SNAPSHOT OFF, RESTRICTED_USER ; Notice that several database options are set in particular ways to ensure we get meaningful and reproducible results from the DMVs.  In particular, the options to auto-create and update statistics are disabled.  There are also three stored procedures, the first of which creates a test table (which may or may not be partitioned).  The table is pretty much the same one we used yesterday: The table has 100 rows, and both the key_col and data columns contain the same values – the integers from 1 to 100 inclusive.  The table is a heap, with a non-clustered primary key on key_col, and a non-clustered non-unique index on the data column.  The only reason I have used a heap here, rather than a clustered table, is so I can demonstrate a seek on a heap later on.  The table has an extra column (not shown because I am too lazy to update the diagram from yesterday) called padding – a CHAR(100) column that just contains 100 spaces in every row.  It’s just there to discourage SQL Server from choosing table scan over an index + RID lookup in one of the tests. The first stored procedure is called ResetTest: CREATE PROCEDURE dbo.ResetTest @Partitioned BIT = 'false' AS BEGIN SET NOCOUNT ON ; IF OBJECT_ID(N'dbo.Example', N'U') IS NOT NULL BEGIN DROP TABLE dbo.Example; END ; -- Test table is a heap -- Non-clustered primary key on 'key_col' CREATE TABLE dbo.Example ( key_col INTEGER NOT NULL, data INTEGER NOT NULL, padding CHAR(100) NOT NULL DEFAULT SPACE(100), CONSTRAINT [PK dbo.Example key_col] PRIMARY KEY NONCLUSTERED (key_col) ) ; IF @Partitioned = 'true' BEGIN -- Enterprise, Trial, or Developer -- required for partitioning tests IF SERVERPROPERTY('EngineEdition') = 3 BEGIN EXECUTE (' DROP TABLE dbo.Example ; IF EXISTS ( SELECT 1 FROM sys.partition_schemes WHERE name = N''PS'' ) DROP PARTITION SCHEME PS ; IF EXISTS ( SELECT 1 FROM sys.partition_functions WHERE name = N''PF'' ) DROP PARTITION FUNCTION PF ; CREATE PARTITION FUNCTION PF (INTEGER) AS RANGE RIGHT FOR VALUES (20, 40, 60, 80, 100) ; CREATE PARTITION SCHEME PS AS PARTITION PF ALL TO ([PRIMARY]) ; CREATE TABLE dbo.Example ( key_col INTEGER NOT NULL, data INTEGER NOT NULL, padding CHAR(100) NOT NULL DEFAULT SPACE(100), CONSTRAINT [PK dbo.Example key_col] PRIMARY KEY NONCLUSTERED (key_col) ) ON PS (key_col); '); END ELSE BEGIN RAISERROR('Invalid SKU for partition test', 16, 1); RETURN; END; END ; -- Non-unique non-clustered index on the 'data' column CREATE NONCLUSTERED INDEX [IX dbo.Example data] ON dbo.Example (data) ; -- Add 100 rows INSERT dbo.Example WITH (TABLOCKX) ( key_col, data ) SELECT key_col = V.number, data = V.number FROM master.dbo.spt_values AS V WHERE V.[type] = N'P' AND V.number BETWEEN 1 AND 100 ; END; GO The second stored procedure, ShowStats, displays information from the Index Usage Stats and Index Operational Stats DMVs: CREATE PROCEDURE dbo.ShowStats @Partitioned BIT = 'false' AS BEGIN -- Index Usage Stats DMV (QE) SELECT index_name = ISNULL(I.name, I.type_desc), scans = IUS.user_scans, seeks = IUS.user_seeks, lookups = IUS.user_lookups FROM sys.dm_db_index_usage_stats AS IUS JOIN sys.indexes AS I ON I.object_id = IUS.object_id AND I.index_id = IUS.index_id WHERE IUS.database_id = DB_ID(N'ScansAndSeeks') AND IUS.object_id = OBJECT_ID(N'dbo.Example', N'U') ORDER BY I.index_id ; -- Index Operational Stats DMV (SE) IF @Partitioned = 'true' SELECT index_name = ISNULL(I.name, I.type_desc), partitions = COUNT(IOS.partition_number), range_scans = SUM(IOS.range_scan_count), single_lookups = SUM(IOS.singleton_lookup_count) FROM sys.dm_db_index_operational_stats ( DB_ID(N'ScansAndSeeks'), OBJECT_ID(N'dbo.Example', N'U'), NULL, NULL ) AS IOS JOIN sys.indexes AS I ON I.object_id = IOS.object_id AND I.index_id = IOS.index_id GROUP BY I.index_id, -- Key I.name, I.type_desc ORDER BY I.index_id; ELSE SELECT index_name = ISNULL(I.name, I.type_desc), range_scans = SUM(IOS.range_scan_count), single_lookups = SUM(IOS.singleton_lookup_count) FROM sys.dm_db_index_operational_stats ( DB_ID(N'ScansAndSeeks'), OBJECT_ID(N'dbo.Example', N'U'), NULL, NULL ) AS IOS JOIN sys.indexes AS I ON I.object_id = IOS.object_id AND I.index_id = IOS.index_id GROUP BY I.index_id, -- Key I.name, I.type_desc ORDER BY I.index_id; END; The final stored procedure, RunTest, executes a query written against the example table: CREATE PROCEDURE dbo.RunTest @SQL VARCHAR(8000), @Partitioned BIT = 'false' AS BEGIN -- No execution plan yet SET STATISTICS XML OFF ; -- Reset the test environment EXECUTE dbo.ResetTest @Partitioned ; -- Previous call will throw an error if a partitioned -- test was requested, but SKU does not support it IF @@ERROR = 0 BEGIN -- IO statistics and plan on SET STATISTICS XML, IO ON ; -- Test statement EXECUTE (@SQL) ; -- Plan and IO statistics off SET STATISTICS XML, IO OFF ; EXECUTE dbo.ShowStats @Partitioned; END; END; The Tests The first test is a simple scan of the heap table: EXECUTE dbo.RunTest @SQL = 'SELECT * FROM Example'; The top result set comes from the Index Usage Stats DMV, so it is the Query Executor’s (QE) view.  The lower result is from Index Operational Stats, which shows statistics derived from the actions taken by the Storage Engine (SE).  We see that QE performed 1 scan operation on the heap, and SE performed a single range scan.  Let’s try a single-value equality seek on a unique index next: EXECUTE dbo.RunTest @SQL = 'SELECT key_col FROM Example WHERE key_col = 32'; This time we see a single seek on the non-clustered primary key from QE, and one singleton lookup on the same index by the SE.  Now for a single-value seek on the non-unique non-clustered index: EXECUTE dbo.RunTest @SQL = 'SELECT data FROM Example WHERE data = 32'; QE shows a single seek on the non-clustered non-unique index, but SE shows a single range scan on that index – not the singleton lookup we saw in the previous test.  That makes sense because we know that only a single-value seek into a unique index is a singleton seek.  A single-value seek into a non-unique index might retrieve any number of rows, if you think about it.  The next query is equivalent to the IN list example seen in the first post in this series, but it is written using OR (just for variety, you understand): EXECUTE dbo.RunTest @SQL = 'SELECT data FROM Example WHERE data = 32 OR data = 33'; The plan looks the same, and there’s no difference in the stats recorded by QE, but the SE shows two range scans.  Again, these are range scans because we are looking for two values in the data column, which is covered by a non-unique index.  I’ve added a snippet from the Properties window to show that the query plan does show two seek predicates, not just one.  Now let’s rewrite the query using BETWEEN: EXECUTE dbo.RunTest @SQL = 'SELECT data FROM Example WHERE data BETWEEN 32 AND 33'; Notice the seek operator only has one predicate now – it’s just a single range scan from 32 to 33 in the index – as the SE output shows.  For the next test, we will look up four values in the key_col column: EXECUTE dbo.RunTest @SQL = 'SELECT key_col FROM Example WHERE key_col IN (2,4,6,8)'; Just a single seek on the PK from the Query Executor, but four singleton lookups reported by the Storage Engine – and four seek predicates in the Properties window.  On to a more complex example: EXECUTE dbo.RunTest @SQL = 'SELECT * FROM Example WITH (INDEX([PK dbo.Example key_col])) WHERE key_col BETWEEN 1 AND 8'; This time we are forcing use of the non-clustered primary key to return eight rows.  The index is not covering for this query, so the query plan includes an RID lookup into the heap to fetch the data and padding columns.  The QE reports a seek on the PK and a lookup on the heap.  The SE reports a single range scan on the PK (to find key_col values between 1 and 8), and eight singleton lookups on the heap.  Remember that a bookmark lookup (RID or Key) is a seek to a single value in a ‘unique index’ – it finds a row in the heap or cluster from a unique RID or clustering key – so that’s why lookups are always singleton lookups, not range scans. Our next example shows what happens when a query plan operator is not executed at all: EXECUTE dbo.RunTest @SQL = 'SELECT key_col FROM Example WHERE key_col = 8 AND @@TRANCOUNT < 0'; The Filter has a start-up predicate which is always false (if your @@TRANCOUNT is less than zero, call CSS immediately).  The index seek is never executed, but QE still records a single seek against the PK because the operator appears once in an executed plan.  The SE output shows no activity at all.  This next example is 2008 and above only, I’m afraid: EXECUTE dbo.RunTest @SQL = 'SELECT * FROM Example WHERE key_col BETWEEN 1 AND 30', @Partitioned = 'true'; This is the first example to use a partitioned table.  QE reports a single seek on the heap (yes – a seek on a heap), and the SE reports two range scans on the heap.  SQL Server knows (from the partitioning definition) that it only needs to look at partitions 1 and 2 to find all the rows where key_col is between 1 and 30 – the engine seeks to find the two partitions, and performs a range scan seek on each partition. The final example for today is another seek on a heap – try to work out the output of the query before running it! EXECUTE dbo.RunTest @SQL = 'SELECT TOP (2) WITH TIES * FROM Example WHERE key_col BETWEEN 1 AND 50 ORDER BY $PARTITION.PF(key_col) DESC', @Partitioned = 'true'; Notice the lack of an explicit Sort operator in the query plan to enforce the ORDER BY clause, and the backward range scan. © 2011 Paul White email: [email protected] twitter: @SQL_Kiwi

    Read the article

  • Day 4 - Game Sprites In Action

    - by dapostolov
    Yesterday I drew an image on the screen. Most exciting, but ... I spent more time blogging about it then actual coding. So this next little while I'm going to streamline my game and research and simply post key notes. Quick notes on the last session: The most important thing I wanted to point out were the following methods:           spriteBatch.Begin(SpriteBlendMode.AlphaBlend);           spriteBatch.Draw(sprite, position, Color.White);           spriteBatch.End(); The spriteBatch object is used to draw Textures and a 2D texture is called a Sprite A texture is generally an image, which is called an Asset in XNA The Draw Method in the Game1.cs is looped (until exit) and utilises the spriteBatch object to draw a Scene To begin drawing a Scene you call the Begin Method. To end a Scene you call the End Method. And to place an image on the Scene you call the Draw method. The most simple implementation of the draw method is:           spriteBatch.Draw(sprite, position, Color.White); 1) sprite - the 2D texture you loaded to draw 2) position - the 2d vector, a set of x & y coordinates 3) Color.White - the tint to apply to the texture, in this case, white light = nothing, nada, no tint. Game Sprites In Action! Today, I played around with Draw methods to get comfortable with their "quirks". The following is an example of the above draw method, but with more parameters available for us to use. Let's investigate!             spriteBatch.Draw(sprite, position2, null, Color.White, MathHelper.ToRadians(45.0f), new Vector2(sprite.Width / 2, sprite.Height / 2), 1.0F, SpriteEffects.None, 0.0F); The parameters (in order): 1) sprite  the texture to display 2) position2 the position on the screen / scene this can also be a rectangle 3) null the portion of the image to display within an image null = display full image this is generally used for animation strips / grids (more on this below) 4) Color.White Texture tinting White = no tint 5) MathHelper.ToRadians(45.0f) rotation of the object, in this case 45 degrees rotates from the set plotting point. 6) new Vector(0,0) the plotting point in this case the top left corner the image will rotate from the top left of the texture in the code above, the point is set to the middle of the image. 7) 1.0f Image scaling (1x) 8) SpriteEffects.None you can flip the image horizontally or vertically 9) 0.0f The z index of the image. 0 = closer, 1 behind? And playing around with different combinations I was able to come up with the following whacky display:   Checking off Yesterdays Intention List: learn game development terminology (in progress) - We learned sprite, scene, texture, and asset. how to place and position (rotate) a static image on the screen (completed) - The thing to note was, it's was in radians and I found a cool helper method to convert degrees into radians. Also, the image rotates from it's specified point. how to layer static images on the screen (completed) - I couldn't seem to get the zIndex working, but one things for sure, the order you draw the image in also determines how it is rendered on the screen. understand image scaling (in progress) - I'm not sure I have this fully covered, but for the most part plug a number in the scaling field and the image grows / shrinks accordingly. can we reuse images? (completed) - yes, I loaded one image and plotted the bugger all over the screen. understand how framerate is handled in XNA (in progress) - I hacked together some code to display the framerate each second. A framerate of 60 appears to be the standard. Interesting to note, the GameTime object does provide you with some cool timing capabilities, such as...is the game running slow? Need to investigate this down the road. how to display text , basic shapes, and colors on the screen (in progress) - i got text rendered on the screen, and i understand containing rectangles. However, I didn't display "shapes" & "colors" how to interact with an image (collision of user input?) (todo) how to animate an image and understand basic animation techniques (in progress) - I was able to create a stripe animation of numbers ranging from 1 - 4, each block was 40 x 40 pixles for a total stripe size of 160 x 40. Using the portion (source Rectangle) parameter, i limited this display to each section at varying intervals. It was interesting to note my first implementation animated at rocket speed. I then tried to create a smoother animation by limiting the redraw capacity, which seemed to work. I guess a little more research will have to be put into this for animating characters / scenes. how to detect colliding images or screen edges (todo) - but the rectangle object can detect collisions I believe. how to manipulate the image, lets say colors, stretching (in progress) - I haven't figured out how to modify a specific color to be another color, but the tinting parameter definately could be used. As for stretching, use the rectangle object as the positioning and the image will stretch to fit! how to focus on a segment of an image...like only displaying a frame on a film reel (completed) - as per basic animation techniques what's the best way to manage images (compression, storage, location, prevent artwork theft, etc.) (todo) Tomorrows Intention Tomorrow I am going to take a stab at rendering a game menu and from there I'm going to investigate how I can improve upon the code and techniques. Intention List: Render a menu, fancy or not Show the mouse cursor Hook up click event A basic animation of somesort Investigate image / menu techniques D.

    Read the article

  • Unexpected behaviour with glFramebufferTexture1D

    - by Roshan
    I am using render to texture concept with glFramebufferTexture1D. I am drawing a cube on non-default FBO with all the vertices as -1,1 (maximum) in X Y Z direction. Now i am setting viewport to X while rendering on non default FBO. My background is blue with white color of cube. For default FBO, i have created 1-D texture and attached this texture to above FBO with color attachment. I am setting width of texture equal to width*height of above FBO view-port. Now, when i render this texture to on another cube, i can see continuous white color on start or end of each face of the cube. That means part of the face is white and rest is blue. I am not sure whether this behavior is correct or not. I expect all the texels should be white as i am using -1 and 1 coordinates for cube rendered on non-default FBO. code: #define WIDTH 3 #define HEIGHT 3 GLfloat vertices8[]={ 1.0f,1.0f,1.0f, -1.0f,1.0f,1.0f, -1.0f,-1.0f,1.0f, 1.0f,-1.0f,1.0f,//face 1 1.0f,-1.0f,-1.0f, -1.0f,-1.0f,-1.0f, -1.0f,1.0f,-1.0f, 1.0f,1.0f,-1.0f,//face 2 1.0f,1.0f,1.0f, 1.0f,-1.0f,1.0f, 1.0f,-1.0f,-1.0f, 1.0f,1.0f,-1.0f,//face 3 -1.0f,1.0f,1.0f, -1.0f,1.0f,-1.0f, -1.0f,-1.0f,-1.0f, -1.0f,-1.0f,1.0f,//face 4 1.0f,1.0f,1.0f, 1.0f,1.0f,-1.0f, -1.0f,1.0f,-1.0f, -1.0f,1.0f,1.0f,//face 5 -1.0f,-1.0f,1.0f, -1.0f,-1.0f,-1.0f, 1.0f,-1.0f,-1.0f, 1.0f,-1.0f,1.0f//face 6 }; GLfloat vertices[]= { 0.5f,0.5f,0.5f, -0.5f,0.5f,0.5f, -0.5f,-0.5f,0.5f, 0.5f,-0.5f,0.5f,//face 1 0.5f,-0.5f,-0.5f, -0.5f,-0.5f,-0.5f, -0.5f,0.5f,-0.5f, 0.5f,0.5f,-0.5f,//face 2 0.5f,0.5f,0.5f, 0.5f,-0.5f,0.5f, 0.5f,-0.5f,-0.5f, 0.5f,0.5f,-0.5f,//face 3 -0.5f,0.5f,0.5f, -0.5f,0.5f,-0.5f, -0.5f,-0.5f,-0.5f, -0.5f,-0.5f,0.5f,//face 4 0.5f,0.5f,0.5f, 0.5f,0.5f,-0.5f, -0.5f,0.5f,-0.5f, -0.5f,0.5f,0.5f,//face 5 -0.5f,-0.5f,0.5f, -0.5f,-0.5f,-0.5f, 0.5f,-0.5f,-0.5f, 0.5f,-0.5f,0.5f//face 6 }; GLuint indices[] = { 0, 2, 1, 0, 3, 2, 4, 5, 6, 4, 6, 7, 8, 9, 10, 8, 10, 11, 12, 15, 14, 12, 14, 13, 16, 17, 18, 16, 18, 19, 20, 23, 22, 20, 22, 21 }; GLfloat texcoord[] = { 0.0, 0.0, 1.0, 0.0, 1.0, 1.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 1.0, 1.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 1.0, 1.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 1.0, 1.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 1.0, 1.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 1.0, 1.0, 0.0, 1.0 }; glGenTextures(1, &id1); glBindTexture(GL_TEXTURE_1D, id1); glGenFramebuffers(1, &Fboid); glTexParameterf(GL_TEXTURE_1D, GL_TEXTURE_MIN_FILTER, GL_NEAREST); glTexParameterf(GL_TEXTURE_1D, GL_TEXTURE_MAG_FILTER, GL_NEAREST); glTexParameterf(GL_TEXTURE_1D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE); glTexImage1D(GL_TEXTURE_1D, 0, GL_RGBA, WIDTH*HEIGHT , 0, GL_RGBA, GL_UNSIGNED_BYTE,0); glBindFramebuffer(GL_FRAMEBUFFER, Fboid); glFramebufferTexture1D(GL_DRAW_FRAMEBUFFER,GL_COLOR_ATTACHMENT0,GL_TEXTURE_1D,id1,0); draw_cube(); glBindFramebuffer(GL_FRAMEBUFFER, 0); draw(); } draw_cube() { glViewport(0, 0, WIDTH, HEIGHT); glClearColor(0.0f, 0.0f, 0.5f, 1.0f); glClear(GL_COLOR_BUFFER_BIT); glEnableVertexAttribArray(glGetAttribLocation(temp.psId,"position")); glVertexAttribPointer(glGetAttribLocation(temp.psId,"position"), 3, GL_FLOAT, GL_FALSE, 0,vertices8); glDrawArrays (GL_TRIANGLE_FAN, 0, 24); } draw() { glClearColor(1.0f, 0.0f, 0.0f, 1.0f); glClearDepth(1.0f); glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT); glEnableVertexAttribArray(glGetAttribLocation(shader_data.psId,"tk_position")); glVertexAttribPointer(glGetAttribLocation(shader_data.psId,"tk_position"), 3, GL_FLOAT, GL_FALSE, 0,vertices); nResult = GL_ERROR_CHECK((GL_NO_ERROR, "glVertexAttribPointer(position, 3, GL_FLOAT, GL_FALSE, 0,vertices);")); glEnableVertexAttribArray(glGetAttribLocation(shader_data.psId,"inputtexcoord")); glVertexAttribPointer(glGetAttribLocation(shader_data.psId,"inputtexcoord"), 2, GL_FLOAT, GL_FALSE, 0,texcoord); glBindTexture(*target11, id1); glDrawElements ( GL_TRIANGLES, 36,GL_UNSIGNED_INT, indices ); when i change WIDTH=HEIGHT=2, and call a glreadpixels with height, width equal to 4 in draw_cube() i can see first 2 pixels with white color, next two with blue(glclearcolor), next two white and then blue and so on.. Now when i change width parameter in glTeximage1D to 16 then ideally i should see alternate patches of white and blue right? But its not the case here. why so?

    Read the article

  • How to change all selected chars to _ in Vim

    - by Kev
    I try to draw a class diagram using Vim. I fill the editor window with white-spaces. Type :match SpellBad /\s/ to highlight all the white-spaces. Ctrl+Q to select vertical white-spaces. Ctrl+I to insert Bar(|) and then Esc ........................... v+l +... + l to select horizontal white-spaces But I don't know how to change all selected horizontal white-spaces to underscore(_). I have to hit _ serval times. When comes to long horizontal line, it's bad. ___________ ___________ | | | | | BaseClass |/__________| Client | |___________|\ |___________| /_\ | |____________________________________ | | | _____|_____ _____|_____ _____|_____ | | | | | | | SubClass1 | | SubClass2 | | SubClass3 | |___________| |___________| |¦¦¦¦¦¦¦¦¦¦¦| I want a quick method to do this. Select it - Change it - Done! Maybe map F6 to do it. Thanks!

    Read the article

  • Improving Partitioned Table Join Performance

    - by Paul White
    The query optimizer does not always choose an optimal strategy when joining partitioned tables. This post looks at an example, showing how a manual rewrite of the query can almost double performance, while reducing the memory grant to almost nothing. Test Data The two tables in this example use a common partitioning partition scheme. The partition function uses 41 equal-size partitions: CREATE PARTITION FUNCTION PFT (integer) AS RANGE RIGHT FOR VALUES ( 125000, 250000, 375000, 500000, 625000, 750000, 875000, 1000000, 1125000, 1250000, 1375000, 1500000, 1625000, 1750000, 1875000, 2000000, 2125000, 2250000, 2375000, 2500000, 2625000, 2750000, 2875000, 3000000, 3125000, 3250000, 3375000, 3500000, 3625000, 3750000, 3875000, 4000000, 4125000, 4250000, 4375000, 4500000, 4625000, 4750000, 4875000, 5000000 ); GO CREATE PARTITION SCHEME PST AS PARTITION PFT ALL TO ([PRIMARY]); There two tables are: CREATE TABLE dbo.T1 ( TID integer NOT NULL IDENTITY(0,1), Column1 integer NOT NULL, Padding binary(100) NOT NULL DEFAULT 0x,   CONSTRAINT PK_T1 PRIMARY KEY CLUSTERED (TID) ON PST (TID) );   CREATE TABLE dbo.T2 ( TID integer NOT NULL, Column1 integer NOT NULL, Padding binary(100) NOT NULL DEFAULT 0x,   CONSTRAINT PK_T2 PRIMARY KEY CLUSTERED (TID, Column1) ON PST (TID) ); The next script loads 5 million rows into T1 with a pseudo-random value between 1 and 5 for Column1. The table is partitioned on the IDENTITY column TID: INSERT dbo.T1 WITH (TABLOCKX) (Column1) SELECT (ABS(CHECKSUM(NEWID())) % 5) + 1 FROM dbo.Numbers AS N WHERE n BETWEEN 1 AND 5000000; In case you don’t already have an auxiliary table of numbers lying around, here’s a script to create one with 10 million rows: CREATE TABLE dbo.Numbers (n bigint PRIMARY KEY);   WITH L0 AS(SELECT 1 AS c UNION ALL SELECT 1), L1 AS(SELECT 1 AS c FROM L0 AS A CROSS JOIN L0 AS B), L2 AS(SELECT 1 AS c FROM L1 AS A CROSS JOIN L1 AS B), L3 AS(SELECT 1 AS c FROM L2 AS A CROSS JOIN L2 AS B), L4 AS(SELECT 1 AS c FROM L3 AS A CROSS JOIN L3 AS B), L5 AS(SELECT 1 AS c FROM L4 AS A CROSS JOIN L4 AS B), Nums AS(SELECT ROW_NUMBER() OVER (ORDER BY (SELECT NULL)) AS n FROM L5) INSERT dbo.Numbers WITH (TABLOCKX) SELECT TOP (10000000) n FROM Nums ORDER BY n OPTION (MAXDOP 1); Table T1 contains data like this: Next we load data into table T2. The relationship between the two tables is that table 2 contains ‘n’ rows for each row in table 1, where ‘n’ is determined by the value in Column1 of table T1. There is nothing particularly special about the data or distribution, by the way. INSERT dbo.T2 WITH (TABLOCKX) (TID, Column1) SELECT T.TID, N.n FROM dbo.T1 AS T JOIN dbo.Numbers AS N ON N.n >= 1 AND N.n <= T.Column1; Table T2 ends up containing about 15 million rows: The primary key for table T2 is a combination of TID and Column1. The data is partitioned according to the value in column TID alone. Partition Distribution The following query shows the number of rows in each partition of table T1: SELECT PartitionID = CA1.P, NumRows = COUNT_BIG(*) FROM dbo.T1 AS T CROSS APPLY (VALUES ($PARTITION.PFT(TID))) AS CA1 (P) GROUP BY CA1.P ORDER BY CA1.P; There are 40 partitions containing 125,000 rows (40 * 125k = 5m rows). The rightmost partition remains empty. The next query shows the distribution for table 2: SELECT PartitionID = CA1.P, NumRows = COUNT_BIG(*) FROM dbo.T2 AS T CROSS APPLY (VALUES ($PARTITION.PFT(TID))) AS CA1 (P) GROUP BY CA1.P ORDER BY CA1.P; There are roughly 375,000 rows in each partition (the rightmost partition is also empty): Ok, that’s the test data done. Test Query and Execution Plan The task is to count the rows resulting from joining tables 1 and 2 on the TID column: SET STATISTICS IO ON; DECLARE @s datetime2 = SYSUTCDATETIME();   SELECT COUNT_BIG(*) FROM dbo.T1 AS T1 JOIN dbo.T2 AS T2 ON T2.TID = T1.TID;   SELECT DATEDIFF(Millisecond, @s, SYSUTCDATETIME()); SET STATISTICS IO OFF; The optimizer chooses a plan using parallel hash join, and partial aggregation: The Plan Explorer plan tree view shows accurate cardinality estimates and an even distribution of rows across threads (click to enlarge the image): With a warm data cache, the STATISTICS IO output shows that no physical I/O was needed, and all 41 partitions were touched: Running the query without actual execution plan or STATISTICS IO information for maximum performance, the query returns in around 2600ms. Execution Plan Analysis The first step toward improving on the execution plan produced by the query optimizer is to understand how it works, at least in outline. The two parallel Clustered Index Scans use multiple threads to read rows from tables T1 and T2. Parallel scan uses a demand-based scheme where threads are given page(s) to scan from the table as needed. This arrangement has certain important advantages, but does result in an unpredictable distribution of rows amongst threads. The point is that multiple threads cooperate to scan the whole table, but it is impossible to predict which rows end up on which threads. For correct results from the parallel hash join, the execution plan has to ensure that rows from T1 and T2 that might join are processed on the same thread. For example, if a row from T1 with join key value ‘1234’ is placed in thread 5’s hash table, the execution plan must guarantee that any rows from T2 that also have join key value ‘1234’ probe thread 5’s hash table for matches. The way this guarantee is enforced in this parallel hash join plan is by repartitioning rows to threads after each parallel scan. The two repartitioning exchanges route rows to threads using a hash function over the hash join keys. The two repartitioning exchanges use the same hash function so rows from T1 and T2 with the same join key must end up on the same hash join thread. Expensive Exchanges This business of repartitioning rows between threads can be very expensive, especially if a large number of rows is involved. The execution plan selected by the optimizer moves 5 million rows through one repartitioning exchange and around 15 million across the other. As a first step toward removing these exchanges, consider the execution plan selected by the optimizer if we join just one partition from each table, disallowing parallelism: SELECT COUNT_BIG(*) FROM dbo.T1 AS T1 JOIN dbo.T2 AS T2 ON T2.TID = T1.TID WHERE $PARTITION.PFT(T1.TID) = 1 AND $PARTITION.PFT(T2.TID) = 1 OPTION (MAXDOP 1); The optimizer has chosen a (one-to-many) merge join instead of a hash join. The single-partition query completes in around 100ms. If everything scaled linearly, we would expect that extending this strategy to all 40 populated partitions would result in an execution time around 4000ms. Using parallelism could reduce that further, perhaps to be competitive with the parallel hash join chosen by the optimizer. This raises a question. If the most efficient way to join one partition from each of the tables is to use a merge join, why does the optimizer not choose a merge join for the full query? Forcing a Merge Join Let’s force the optimizer to use a merge join on the test query using a hint: SELECT COUNT_BIG(*) FROM dbo.T1 AS T1 JOIN dbo.T2 AS T2 ON T2.TID = T1.TID OPTION (MERGE JOIN); This is the execution plan selected by the optimizer: This plan results in the same number of logical reads reported previously, but instead of 2600ms the query takes 5000ms. The natural explanation for this drop in performance is that the merge join plan is only using a single thread, whereas the parallel hash join plan could use multiple threads. Parallel Merge Join We can get a parallel merge join plan using the same query hint as before, and adding trace flag 8649: SELECT COUNT_BIG(*) FROM dbo.T1 AS T1 JOIN dbo.T2 AS T2 ON T2.TID = T1.TID OPTION (MERGE JOIN, QUERYTRACEON 8649); The execution plan is: This looks promising. It uses a similar strategy to distribute work across threads as seen for the parallel hash join. In practice though, performance is disappointing. On a typical run, the parallel merge plan runs for around 8400ms; slower than the single-threaded merge join plan (5000ms) and much worse than the 2600ms for the parallel hash join. We seem to be going backwards! The logical reads for the parallel merge are still exactly the same as before, with no physical IOs. The cardinality estimates and thread distribution are also still very good (click to enlarge): A big clue to the reason for the poor performance is shown in the wait statistics (captured by Plan Explorer Pro): CXPACKET waits require careful interpretation, and are most often benign, but in this case excessive waiting occurs at the repartitioning exchanges. Unlike the parallel hash join, the repartitioning exchanges in this plan are order-preserving ‘merging’ exchanges (because merge join requires ordered inputs): Parallelism works best when threads can just grab any available unit of work and get on with processing it. Preserving order introduces inter-thread dependencies that can easily lead to significant waits occurring. In extreme cases, these dependencies can result in an intra-query deadlock, though the details of that will have to wait for another time to explore in detail. The potential for waits and deadlocks leads the query optimizer to cost parallel merge join relatively highly, especially as the degree of parallelism (DOP) increases. This high costing resulted in the optimizer choosing a serial merge join rather than parallel in this case. The test results certainly confirm its reasoning. Collocated Joins In SQL Server 2008 and later, the optimizer has another available strategy when joining tables that share a common partition scheme. This strategy is a collocated join, also known as as a per-partition join. It can be applied in both serial and parallel execution plans, though it is limited to 2-way joins in the current optimizer. Whether the optimizer chooses a collocated join or not depends on cost estimation. The primary benefits of a collocated join are that it eliminates an exchange and requires less memory, as we will see next. Costing and Plan Selection The query optimizer did consider a collocated join for our original query, but it was rejected on cost grounds. The parallel hash join with repartitioning exchanges appeared to be a cheaper option. There is no query hint to force a collocated join, so we have to mess with the costing framework to produce one for our test query. Pretending that IOs cost 50 times more than usual is enough to convince the optimizer to use collocated join with our test query: -- Pretend IOs are 50x cost temporarily DBCC SETIOWEIGHT(50);   -- Co-located hash join SELECT COUNT_BIG(*) FROM dbo.T1 AS T1 JOIN dbo.T2 AS T2 ON T2.TID = T1.TID OPTION (RECOMPILE);   -- Reset IO costing DBCC SETIOWEIGHT(1); Collocated Join Plan The estimated execution plan for the collocated join is: The Constant Scan contains one row for each partition of the shared partitioning scheme, from 1 to 41. The hash repartitioning exchanges seen previously are replaced by a single Distribute Streams exchange using Demand partitioning. Demand partitioning means that the next partition id is given to the next parallel thread that asks for one. My test machine has eight logical processors, and all are available for SQL Server to use. As a result, there are eight threads in the single parallel branch in this plan, each processing one partition from each table at a time. Once a thread finishes processing a partition, it grabs a new partition number from the Distribute Streams exchange…and so on until all partitions have been processed. It is important to understand that the parallel scans in this plan are different from the parallel hash join plan. Although the scans have the same parallelism icon, tables T1 and T2 are not being co-operatively scanned by multiple threads in the same way. Each thread reads a single partition of T1 and performs a hash match join with the same partition from table T2. The properties of the two Clustered Index Scans show a Seek Predicate (unusual for a scan!) limiting the rows to a single partition: The crucial point is that the join between T1 and T2 is on TID, and TID is the partitioning column for both tables. A thread that processes partition ‘n’ is guaranteed to see all rows that can possibly join on TID for that partition. In addition, no other thread will see rows from that partition, so this removes the need for repartitioning exchanges. CPU and Memory Efficiency Improvements The collocated join has removed two expensive repartitioning exchanges and added a single exchange processing 41 rows (one for each partition id). Remember, the parallel hash join plan exchanges had to process 5 million and 15 million rows. The amount of processor time spent on exchanges will be much lower in the collocated join plan. In addition, the collocated join plan has a maximum of 8 threads processing single partitions at any one time. The 41 partitions will all be processed eventually, but a new partition is not started until a thread asks for it. Threads can reuse hash table memory for the new partition. The parallel hash join plan also had 8 hash tables, but with all 5,000,000 build rows loaded at the same time. The collocated plan needs memory for only 8 * 125,000 = 1,000,000 rows at any one time. Collocated Hash Join Performance The collated join plan has disappointing performance in this case. The query runs for around 25,300ms despite the same IO statistics as usual. This is much the worst result so far, so what went wrong? It turns out that cardinality estimation for the single partition scans of table T1 is slightly low. The properties of the Clustered Index Scan of T1 (graphic immediately above) show the estimation was for 121,951 rows. This is a small shortfall compared with the 125,000 rows actually encountered, but it was enough to cause the hash join to spill to physical tempdb: A level 1 spill doesn’t sound too bad, until you realize that the spill to tempdb probably occurs for each of the 41 partitions. As a side note, the cardinality estimation error is a little surprising because the system tables accurately show there are 125,000 rows in every partition of T1. Unfortunately, the optimizer uses regular column and index statistics to derive cardinality estimates here rather than system table information (e.g. sys.partitions). Collocated Merge Join We will never know how well the collocated parallel hash join plan might have worked without the cardinality estimation error (and the resulting 41 spills to tempdb) but we do know: Merge join does not require a memory grant; and Merge join was the optimizer’s preferred join option for a single partition join Putting this all together, what we would really like to see is the same collocated join strategy, but using merge join instead of hash join. Unfortunately, the current query optimizer cannot produce a collocated merge join; it only knows how to do collocated hash join. So where does this leave us? CROSS APPLY sys.partitions We can try to write our own collocated join query. We can use sys.partitions to find the partition numbers, and CROSS APPLY to get a count per partition, with a final step to sum the partial counts. The following query implements this idea: SELECT row_count = SUM(Subtotals.cnt) FROM ( -- Partition numbers SELECT p.partition_number FROM sys.partitions AS p WHERE p.[object_id] = OBJECT_ID(N'T1', N'U') AND p.index_id = 1 ) AS P CROSS APPLY ( -- Count per collocated join SELECT cnt = COUNT_BIG(*) FROM dbo.T1 AS T1 JOIN dbo.T2 AS T2 ON T2.TID = T1.TID WHERE $PARTITION.PFT(T1.TID) = p.partition_number AND $PARTITION.PFT(T2.TID) = p.partition_number ) AS SubTotals; The estimated plan is: The cardinality estimates aren’t all that good here, especially the estimate for the scan of the system table underlying the sys.partitions view. Nevertheless, the plan shape is heading toward where we would like to be. Each partition number from the system table results in a per-partition scan of T1 and T2, a one-to-many Merge Join, and a Stream Aggregate to compute the partial counts. The final Stream Aggregate just sums the partial counts. Execution time for this query is around 3,500ms, with the same IO statistics as always. This compares favourably with 5,000ms for the serial plan produced by the optimizer with the OPTION (MERGE JOIN) hint. This is another case of the sum of the parts being less than the whole – summing 41 partial counts from 41 single-partition merge joins is faster than a single merge join and count over all partitions. Even so, this single-threaded collocated merge join is not as quick as the original parallel hash join plan, which executed in 2,600ms. On the positive side, our collocated merge join uses only one logical processor and requires no memory grant. The parallel hash join plan used 16 threads and reserved 569 MB of memory:   Using a Temporary Table Our collocated merge join plan should benefit from parallelism. The reason parallelism is not being used is that the query references a system table. We can work around that by writing the partition numbers to a temporary table (or table variable): SET STATISTICS IO ON; DECLARE @s datetime2 = SYSUTCDATETIME();   CREATE TABLE #P ( partition_number integer PRIMARY KEY);   INSERT #P (partition_number) SELECT p.partition_number FROM sys.partitions AS p WHERE p.[object_id] = OBJECT_ID(N'T1', N'U') AND p.index_id = 1;   SELECT row_count = SUM(Subtotals.cnt) FROM #P AS p CROSS APPLY ( SELECT cnt = COUNT_BIG(*) FROM dbo.T1 AS T1 JOIN dbo.T2 AS T2 ON T2.TID = T1.TID WHERE $PARTITION.PFT(T1.TID) = p.partition_number AND $PARTITION.PFT(T2.TID) = p.partition_number ) AS SubTotals;   DROP TABLE #P;   SELECT DATEDIFF(Millisecond, @s, SYSUTCDATETIME()); SET STATISTICS IO OFF; Using the temporary table adds a few logical reads, but the overall execution time is still around 3500ms, indistinguishable from the same query without the temporary table. The problem is that the query optimizer still doesn’t choose a parallel plan for this query, though the removal of the system table reference means that it could if it chose to: In fact the optimizer did enter the parallel plan phase of query optimization (running search 1 for a second time): Unfortunately, the parallel plan found seemed to be more expensive than the serial plan. This is a crazy result, caused by the optimizer’s cost model not reducing operator CPU costs on the inner side of a nested loops join. Don’t get me started on that, we’ll be here all night. In this plan, everything expensive happens on the inner side of a nested loops join. Without a CPU cost reduction to compensate for the added cost of exchange operators, candidate parallel plans always look more expensive to the optimizer than the equivalent serial plan. Parallel Collocated Merge Join We can produce the desired parallel plan using trace flag 8649 again: SELECT row_count = SUM(Subtotals.cnt) FROM #P AS p CROSS APPLY ( SELECT cnt = COUNT_BIG(*) FROM dbo.T1 AS T1 JOIN dbo.T2 AS T2 ON T2.TID = T1.TID WHERE $PARTITION.PFT(T1.TID) = p.partition_number AND $PARTITION.PFT(T2.TID) = p.partition_number ) AS SubTotals OPTION (QUERYTRACEON 8649); The actual execution plan is: One difference between this plan and the collocated hash join plan is that a Repartition Streams exchange operator is used instead of Distribute Streams. The effect is similar, though not quite identical. The Repartition uses round-robin partitioning, meaning the next partition id is pushed to the next thread in sequence. The Distribute Streams exchange seen earlier used Demand partitioning, meaning the next partition id is pulled across the exchange by the next thread that is ready for more work. There are subtle performance implications for each partitioning option, but going into that would again take us too far off the main point of this post. Performance The important thing is the performance of this parallel collocated merge join – just 1350ms on a typical run. The list below shows all the alternatives from this post (all timings include creation, population, and deletion of the temporary table where appropriate) from quickest to slowest: Collocated parallel merge join: 1350ms Parallel hash join: 2600ms Collocated serial merge join: 3500ms Serial merge join: 5000ms Parallel merge join: 8400ms Collated parallel hash join: 25,300ms (hash spill per partition) The parallel collocated merge join requires no memory grant (aside from a paltry 1.2MB used for exchange buffers). This plan uses 16 threads at DOP 8; but 8 of those are (rather pointlessly) allocated to the parallel scan of the temporary table. These are minor concerns, but it turns out there is a way to address them if it bothers you. Parallel Collocated Merge Join with Demand Partitioning This final tweak replaces the temporary table with a hard-coded list of partition ids (dynamic SQL could be used to generate this query from sys.partitions): SELECT row_count = SUM(Subtotals.cnt) FROM ( VALUES (1),(2),(3),(4),(5),(6),(7),(8),(9),(10), (11),(12),(13),(14),(15),(16),(17),(18),(19),(20), (21),(22),(23),(24),(25),(26),(27),(28),(29),(30), (31),(32),(33),(34),(35),(36),(37),(38),(39),(40),(41) ) AS P (partition_number) CROSS APPLY ( SELECT cnt = COUNT_BIG(*) FROM dbo.T1 AS T1 JOIN dbo.T2 AS T2 ON T2.TID = T1.TID WHERE $PARTITION.PFT(T1.TID) = p.partition_number AND $PARTITION.PFT(T2.TID) = p.partition_number ) AS SubTotals OPTION (QUERYTRACEON 8649); The actual execution plan is: The parallel collocated hash join plan is reproduced below for comparison: The manual rewrite has another advantage that has not been mentioned so far: the partial counts (per partition) can be computed earlier than the partial counts (per thread) in the optimizer’s collocated join plan. The earlier aggregation is performed by the extra Stream Aggregate under the nested loops join. The performance of the parallel collocated merge join is unchanged at around 1350ms. Final Words It is a shame that the current query optimizer does not consider a collocated merge join (Connect item closed as Won’t Fix). The example used in this post showed an improvement in execution time from 2600ms to 1350ms using a modestly-sized data set and limited parallelism. In addition, the memory requirement for the query was almost completely eliminated  – down from 569MB to 1.2MB. The problem with the parallel hash join selected by the optimizer is that it attempts to process the full data set all at once (albeit using eight threads). It requires a large memory grant to hold all 5 million rows from table T1 across the eight hash tables, and does not take advantage of the divide-and-conquer opportunity offered by the common partitioning. The great thing about the collocated join strategies is that each parallel thread works on a single partition from both tables, reading rows, performing the join, and computing a per-partition subtotal, before moving on to a new partition. From a thread’s point of view… If you have trouble visualizing what is happening from just looking at the parallel collocated merge join execution plan, let’s look at it again, but from the point of view of just one thread operating between the two Parallelism (exchange) operators. Our thread picks up a single partition id from the Distribute Streams exchange, and starts a merge join using ordered rows from partition 1 of table T1 and partition 1 of table T2. By definition, this is all happening on a single thread. As rows join, they are added to a (per-partition) count in the Stream Aggregate immediately above the Merge Join. Eventually, either T1 (partition 1) or T2 (partition 1) runs out of rows and the merge join stops. The per-partition count from the aggregate passes on through the Nested Loops join to another Stream Aggregate, which is maintaining a per-thread subtotal. Our same thread now picks up a new partition id from the exchange (say it gets id 9 this time). The count in the per-partition aggregate is reset to zero, and the processing of partition 9 of both tables proceeds just as it did for partition 1, and on the same thread. Each thread picks up a single partition id and processes all the data for that partition, completely independently from other threads working on other partitions. One thread might eventually process partitions (1, 9, 17, 25, 33, 41) while another is concurrently processing partitions (2, 10, 18, 26, 34) and so on for the other six threads at DOP 8. The point is that all 8 threads can execute independently and concurrently, continuing to process new partitions until the wider job (of which the thread has no knowledge!) is done. This divide-and-conquer technique can be much more efficient than simply splitting the entire workload across eight threads all at once. Related Reading Understanding and Using Parallelism in SQL Server Parallel Execution Plans Suck © 2013 Paul White – All Rights Reserved Twitter: @SQL_Kiwi

    Read the article

  • Advanced TSQL Tuning: Why Internals Knowledge Matters

    - by Paul White
    There is much more to query tuning than reducing logical reads and adding covering nonclustered indexes.  Query tuning is not complete as soon as the query returns results quickly in the development or test environments.  In production, your query will compete for memory, CPU, locks, I/O and other resources on the server.  Today’s entry looks at some tuning considerations that are often overlooked, and shows how deep internals knowledge can help you write better TSQL. As always, we’ll need some example data.  In fact, we are going to use three tables today, each of which is structured like this: Each table has 50,000 rows made up of an INTEGER id column and a padding column containing 3,999 characters in every row.  The only difference between the three tables is in the type of the padding column: the first table uses CHAR(3999), the second uses VARCHAR(MAX), and the third uses the deprecated TEXT type.  A script to create a database with the three tables and load the sample data follows: USE master; GO IF DB_ID('SortTest') IS NOT NULL DROP DATABASE SortTest; GO CREATE DATABASE SortTest COLLATE LATIN1_GENERAL_BIN; GO ALTER DATABASE SortTest MODIFY FILE ( NAME = 'SortTest', SIZE = 3GB, MAXSIZE = 3GB ); GO ALTER DATABASE SortTest MODIFY FILE ( NAME = 'SortTest_log', SIZE = 256MB, MAXSIZE = 1GB, FILEGROWTH = 128MB ); GO ALTER DATABASE SortTest SET ALLOW_SNAPSHOT_ISOLATION OFF ; ALTER DATABASE SortTest SET AUTO_CLOSE OFF ; ALTER DATABASE SortTest SET AUTO_CREATE_STATISTICS ON ; ALTER DATABASE SortTest SET AUTO_SHRINK OFF ; ALTER DATABASE SortTest SET AUTO_UPDATE_STATISTICS ON ; ALTER DATABASE SortTest SET AUTO_UPDATE_STATISTICS_ASYNC ON ; ALTER DATABASE SortTest SET PARAMETERIZATION SIMPLE ; ALTER DATABASE SortTest SET READ_COMMITTED_SNAPSHOT OFF ; ALTER DATABASE SortTest SET MULTI_USER ; ALTER DATABASE SortTest SET RECOVERY SIMPLE ; USE SortTest; GO CREATE TABLE dbo.TestCHAR ( id INTEGER IDENTITY (1,1) NOT NULL, padding CHAR(3999) NOT NULL,   CONSTRAINT [PK dbo.TestCHAR (id)] PRIMARY KEY CLUSTERED (id), ) ; CREATE TABLE dbo.TestMAX ( id INTEGER IDENTITY (1,1) NOT NULL, padding VARCHAR(MAX) NOT NULL,   CONSTRAINT [PK dbo.TestMAX (id)] PRIMARY KEY CLUSTERED (id), ) ; CREATE TABLE dbo.TestTEXT ( id INTEGER IDENTITY (1,1) NOT NULL, padding TEXT NOT NULL,   CONSTRAINT [PK dbo.TestTEXT (id)] PRIMARY KEY CLUSTERED (id), ) ; -- ============= -- Load TestCHAR (about 3s) -- ============= INSERT INTO dbo.TestCHAR WITH (TABLOCKX) ( padding ) SELECT padding = REPLICATE(CHAR(65 + (Data.n % 26)), 3999) FROM ( SELECT TOP (50000) n = ROW_NUMBER() OVER (ORDER BY (SELECT 0)) - 1 FROM master.sys.columns C1, master.sys.columns C2, master.sys.columns C3 ORDER BY n ASC ) AS Data ORDER BY Data.n ASC ; -- ============ -- Load TestMAX (about 3s) -- ============ INSERT INTO dbo.TestMAX WITH (TABLOCKX) ( padding ) SELECT CONVERT(VARCHAR(MAX), padding) FROM dbo.TestCHAR ORDER BY id ; -- ============= -- Load TestTEXT (about 5s) -- ============= INSERT INTO dbo.TestTEXT WITH (TABLOCKX) ( padding ) SELECT CONVERT(TEXT, padding) FROM dbo.TestCHAR ORDER BY id ; -- ========== -- Space used -- ========== -- EXECUTE sys.sp_spaceused @objname = 'dbo.TestCHAR'; EXECUTE sys.sp_spaceused @objname = 'dbo.TestMAX'; EXECUTE sys.sp_spaceused @objname = 'dbo.TestTEXT'; ; CHECKPOINT ; That takes around 15 seconds to run, and shows the space allocated to each table in its output: To illustrate the points I want to make today, the example task we are going to set ourselves is to return a random set of 150 rows from each table.  The basic shape of the test query is the same for each of the three test tables: SELECT TOP (150) T.id, T.padding FROM dbo.Test AS T ORDER BY NEWID() OPTION (MAXDOP 1) ; Test 1 – CHAR(3999) Running the template query shown above using the TestCHAR table as the target, we find that the query takes around 5 seconds to return its results.  This seems slow, considering that the table only has 50,000 rows.  Working on the assumption that generating a GUID for each row is a CPU-intensive operation, we might try enabling parallelism to see if that speeds up the response time.  Running the query again (but without the MAXDOP 1 hint) on a machine with eight logical processors, the query now takes 10 seconds to execute – twice as long as when run serially. Rather than attempting further guesses at the cause of the slowness, let’s go back to serial execution and add some monitoring.  The script below monitors STATISTICS IO output and the amount of tempdb used by the test query.  We will also run a Profiler trace to capture any warnings generated during query execution. DECLARE @read BIGINT, @write BIGINT ; SELECT @read = SUM(num_of_bytes_read), @write = SUM(num_of_bytes_written) FROM tempdb.sys.database_files AS DBF JOIN sys.dm_io_virtual_file_stats(2, NULL) AS FS ON FS.file_id = DBF.file_id WHERE DBF.type_desc = 'ROWS' ; SET STATISTICS IO ON ; SELECT TOP (150) TC.id, TC.padding FROM dbo.TestCHAR AS TC ORDER BY NEWID() OPTION (MAXDOP 1) ; SET STATISTICS IO OFF ; SELECT tempdb_read_MB = (SUM(num_of_bytes_read) - @read) / 1024. / 1024., tempdb_write_MB = (SUM(num_of_bytes_written) - @write) / 1024. / 1024., internal_use_MB = ( SELECT internal_objects_alloc_page_count / 128.0 FROM sys.dm_db_task_space_usage WHERE session_id = @@SPID ) FROM tempdb.sys.database_files AS DBF JOIN sys.dm_io_virtual_file_stats(2, NULL) AS FS ON FS.file_id = DBF.file_id WHERE DBF.type_desc = 'ROWS' ; Let’s take a closer look at the statistics and query plan generated from this: Following the flow of the data from right to left, we see the expected 50,000 rows emerging from the Clustered Index Scan, with a total estimated size of around 191MB.  The Compute Scalar adds a column containing a random GUID (generated from the NEWID() function call) for each row.  With this extra column in place, the size of the data arriving at the Sort operator is estimated to be 192MB. Sort is a blocking operator – it has to examine all of the rows on its input before it can produce its first row of output (the last row received might sort first).  This characteristic means that Sort requires a memory grant – memory allocated for the query’s use by SQL Server just before execution starts.  In this case, the Sort is the only memory-consuming operator in the plan, so it has access to the full 243MB (248,696KB) of memory reserved by SQL Server for this query execution. Notice that the memory grant is significantly larger than the expected size of the data to be sorted.  SQL Server uses a number of techniques to speed up sorting, some of which sacrifice size for comparison speed.  Sorts typically require a very large number of comparisons, so this is usually a very effective optimization.  One of the drawbacks is that it is not possible to exactly predict the sort space needed, as it depends on the data itself.  SQL Server takes an educated guess based on data types, sizes, and the number of rows expected, but the algorithm is not perfect. In spite of the large memory grant, the Profiler trace shows a Sort Warning event (indicating that the sort ran out of memory), and the tempdb usage monitor shows that 195MB of tempdb space was used – all of that for system use.  The 195MB represents physical write activity on tempdb, because SQL Server strictly enforces memory grants – a query cannot ‘cheat’ and effectively gain extra memory by spilling to tempdb pages that reside in memory.  Anyway, the key point here is that it takes a while to write 195MB to disk, and this is the main reason that the query takes 5 seconds overall. If you are wondering why using parallelism made the problem worse, consider that eight threads of execution result in eight concurrent partial sorts, each receiving one eighth of the memory grant.  The eight sorts all spilled to tempdb, resulting in inefficiencies as the spilled sorts competed for disk resources.  More importantly, there are specific problems at the point where the eight partial results are combined, but I’ll cover that in a future post. CHAR(3999) Performance Summary: 5 seconds elapsed time 243MB memory grant 195MB tempdb usage 192MB estimated sort set 25,043 logical reads Sort Warning Test 2 – VARCHAR(MAX) We’ll now run exactly the same test (with the additional monitoring) on the table using a VARCHAR(MAX) padding column: DECLARE @read BIGINT, @write BIGINT ; SELECT @read = SUM(num_of_bytes_read), @write = SUM(num_of_bytes_written) FROM tempdb.sys.database_files AS DBF JOIN sys.dm_io_virtual_file_stats(2, NULL) AS FS ON FS.file_id = DBF.file_id WHERE DBF.type_desc = 'ROWS' ; SET STATISTICS IO ON ; SELECT TOP (150) TM.id, TM.padding FROM dbo.TestMAX AS TM ORDER BY NEWID() OPTION (MAXDOP 1) ; SET STATISTICS IO OFF ; SELECT tempdb_read_MB = (SUM(num_of_bytes_read) - @read) / 1024. / 1024., tempdb_write_MB = (SUM(num_of_bytes_written) - @write) / 1024. / 1024., internal_use_MB = ( SELECT internal_objects_alloc_page_count / 128.0 FROM sys.dm_db_task_space_usage WHERE session_id = @@SPID ) FROM tempdb.sys.database_files AS DBF JOIN sys.dm_io_virtual_file_stats(2, NULL) AS FS ON FS.file_id = DBF.file_id WHERE DBF.type_desc = 'ROWS' ; This time the query takes around 8 seconds to complete (3 seconds longer than Test 1).  Notice that the estimated row and data sizes are very slightly larger, and the overall memory grant has also increased very slightly to 245MB.  The most marked difference is in the amount of tempdb space used – this query wrote almost 391MB of sort run data to the physical tempdb file.  Don’t draw any general conclusions about VARCHAR(MAX) versus CHAR from this – I chose the length of the data specifically to expose this edge case.  In most cases, VARCHAR(MAX) performs very similarly to CHAR – I just wanted to make test 2 a bit more exciting. MAX Performance Summary: 8 seconds elapsed time 245MB memory grant 391MB tempdb usage 193MB estimated sort set 25,043 logical reads Sort warning Test 3 – TEXT The same test again, but using the deprecated TEXT data type for the padding column: DECLARE @read BIGINT, @write BIGINT ; SELECT @read = SUM(num_of_bytes_read), @write = SUM(num_of_bytes_written) FROM tempdb.sys.database_files AS DBF JOIN sys.dm_io_virtual_file_stats(2, NULL) AS FS ON FS.file_id = DBF.file_id WHERE DBF.type_desc = 'ROWS' ; SET STATISTICS IO ON ; SELECT TOP (150) TT.id, TT.padding FROM dbo.TestTEXT AS TT ORDER BY NEWID() OPTION (MAXDOP 1, RECOMPILE) ; SET STATISTICS IO OFF ; SELECT tempdb_read_MB = (SUM(num_of_bytes_read) - @read) / 1024. / 1024., tempdb_write_MB = (SUM(num_of_bytes_written) - @write) / 1024. / 1024., internal_use_MB = ( SELECT internal_objects_alloc_page_count / 128.0 FROM sys.dm_db_task_space_usage WHERE session_id = @@SPID ) FROM tempdb.sys.database_files AS DBF JOIN sys.dm_io_virtual_file_stats(2, NULL) AS FS ON FS.file_id = DBF.file_id WHERE DBF.type_desc = 'ROWS' ; This time the query runs in 500ms.  If you look at the metrics we have been checking so far, it’s not hard to understand why: TEXT Performance Summary: 0.5 seconds elapsed time 9MB memory grant 5MB tempdb usage 5MB estimated sort set 207 logical reads 596 LOB logical reads Sort warning SQL Server’s memory grant algorithm still underestimates the memory needed to perform the sorting operation, but the size of the data to sort is so much smaller (5MB versus 193MB previously) that the spilled sort doesn’t matter very much.  Why is the data size so much smaller?  The query still produces the correct results – including the large amount of data held in the padding column – so what magic is being performed here? TEXT versus MAX Storage The answer lies in how columns of the TEXT data type are stored.  By default, TEXT data is stored off-row in separate LOB pages – which explains why this is the first query we have seen that records LOB logical reads in its STATISTICS IO output.  You may recall from my last post that LOB data leaves an in-row pointer to the separate storage structure holding the LOB data. SQL Server can see that the full LOB value is not required by the query plan until results are returned, so instead of passing the full LOB value down the plan from the Clustered Index Scan, it passes the small in-row structure instead.  SQL Server estimates that each row coming from the scan will be 79 bytes long – 11 bytes for row overhead, 4 bytes for the integer id column, and 64 bytes for the LOB pointer (in fact the pointer is rather smaller – usually 16 bytes – but the details of that don’t really matter right now). OK, so this query is much more efficient because it is sorting a very much smaller data set – SQL Server delays retrieving the LOB data itself until after the Sort starts producing its 150 rows.  The question that normally arises at this point is: Why doesn’t SQL Server use the same trick when the padding column is defined as VARCHAR(MAX)? The answer is connected with the fact that if the actual size of the VARCHAR(MAX) data is 8000 bytes or less, it is usually stored in-row in exactly the same way as for a VARCHAR(8000) column – MAX data only moves off-row into LOB storage when it exceeds 8000 bytes.  The default behaviour of the TEXT type is to be stored off-row by default, unless the ‘text in row’ table option is set suitably and there is room on the page.  There is an analogous (but opposite) setting to control the storage of MAX data – the ‘large value types out of row’ table option.  By enabling this option for a table, MAX data will be stored off-row (in a LOB structure) instead of in-row.  SQL Server Books Online has good coverage of both options in the topic In Row Data. The MAXOOR Table The essential difference, then, is that MAX defaults to in-row storage, and TEXT defaults to off-row (LOB) storage.  You might be thinking that we could get the same benefits seen for the TEXT data type by storing the VARCHAR(MAX) values off row – so let’s look at that option now.  This script creates a fourth table, with the VARCHAR(MAX) data stored off-row in LOB pages: CREATE TABLE dbo.TestMAXOOR ( id INTEGER IDENTITY (1,1) NOT NULL, padding VARCHAR(MAX) NOT NULL,   CONSTRAINT [PK dbo.TestMAXOOR (id)] PRIMARY KEY CLUSTERED (id), ) ; EXECUTE sys.sp_tableoption @TableNamePattern = N'dbo.TestMAXOOR', @OptionName = 'large value types out of row', @OptionValue = 'true' ; SELECT large_value_types_out_of_row FROM sys.tables WHERE [schema_id] = SCHEMA_ID(N'dbo') AND name = N'TestMAXOOR' ; INSERT INTO dbo.TestMAXOOR WITH (TABLOCKX) ( padding ) SELECT SPACE(0) FROM dbo.TestCHAR ORDER BY id ; UPDATE TM WITH (TABLOCK) SET padding.WRITE (TC.padding, NULL, NULL) FROM dbo.TestMAXOOR AS TM JOIN dbo.TestCHAR AS TC ON TC.id = TM.id ; EXECUTE sys.sp_spaceused @objname = 'dbo.TestMAXOOR' ; CHECKPOINT ; Test 4 – MAXOOR We can now re-run our test on the MAXOOR (MAX out of row) table: DECLARE @read BIGINT, @write BIGINT ; SELECT @read = SUM(num_of_bytes_read), @write = SUM(num_of_bytes_written) FROM tempdb.sys.database_files AS DBF JOIN sys.dm_io_virtual_file_stats(2, NULL) AS FS ON FS.file_id = DBF.file_id WHERE DBF.type_desc = 'ROWS' ; SET STATISTICS IO ON ; SELECT TOP (150) MO.id, MO.padding FROM dbo.TestMAXOOR AS MO ORDER BY NEWID() OPTION (MAXDOP 1, RECOMPILE) ; SET STATISTICS IO OFF ; SELECT tempdb_read_MB = (SUM(num_of_bytes_read) - @read) / 1024. / 1024., tempdb_write_MB = (SUM(num_of_bytes_written) - @write) / 1024. / 1024., internal_use_MB = ( SELECT internal_objects_alloc_page_count / 128.0 FROM sys.dm_db_task_space_usage WHERE session_id = @@SPID ) FROM tempdb.sys.database_files AS DBF JOIN sys.dm_io_virtual_file_stats(2, NULL) AS FS ON FS.file_id = DBF.file_id WHERE DBF.type_desc = 'ROWS' ; TEXT Performance Summary: 0.3 seconds elapsed time 245MB memory grant 0MB tempdb usage 193MB estimated sort set 207 logical reads 446 LOB logical reads No sort warning The query runs very quickly – slightly faster than Test 3, and without spilling the sort to tempdb (there is no sort warning in the trace, and the monitoring query shows zero tempdb usage by this query).  SQL Server is passing the in-row pointer structure down the plan and only looking up the LOB value on the output side of the sort. The Hidden Problem There is still a huge problem with this query though – it requires a 245MB memory grant.  No wonder the sort doesn’t spill to tempdb now – 245MB is about 20 times more memory than this query actually requires to sort 50,000 records containing LOB data pointers.  Notice that the estimated row and data sizes in the plan are the same as in test 2 (where the MAX data was stored in-row). The optimizer assumes that MAX data is stored in-row, regardless of the sp_tableoption setting ‘large value types out of row’.  Why?  Because this option is dynamic – changing it does not immediately force all MAX data in the table in-row or off-row, only when data is added or actually changed.  SQL Server does not keep statistics to show how much MAX or TEXT data is currently in-row, and how much is stored in LOB pages.  This is an annoying limitation, and one which I hope will be addressed in a future version of the product. So why should we worry about this?  Excessive memory grants reduce concurrency and may result in queries waiting on the RESOURCE_SEMAPHORE wait type while they wait for memory they do not need.  245MB is an awful lot of memory, especially on 32-bit versions where memory grants cannot use AWE-mapped memory.  Even on a 64-bit server with plenty of memory, do you really want a single query to consume 0.25GB of memory unnecessarily?  That’s 32,000 8KB pages that might be put to much better use. The Solution The answer is not to use the TEXT data type for the padding column.  That solution happens to have better performance characteristics for this specific query, but it still results in a spilled sort, and it is hard to recommend the use of a data type which is scheduled for removal.  I hope it is clear to you that the fundamental problem here is that SQL Server sorts the whole set arriving at a Sort operator.  Clearly, it is not efficient to sort the whole table in memory just to return 150 rows in a random order. The TEXT example was more efficient because it dramatically reduced the size of the set that needed to be sorted.  We can do the same thing by selecting 150 unique keys from the table at random (sorting by NEWID() for example) and only then retrieving the large padding column values for just the 150 rows we need.  The following script implements that idea for all four tables: SET STATISTICS IO ON ; WITH TestTable AS ( SELECT * FROM dbo.TestCHAR ), TopKeys AS ( SELECT TOP (150) id FROM TestTable ORDER BY NEWID() ) SELECT T1.id, T1.padding FROM TestTable AS T1 WHERE T1.id = ANY (SELECT id FROM TopKeys) OPTION (MAXDOP 1) ; WITH TestTable AS ( SELECT * FROM dbo.TestMAX ), TopKeys AS ( SELECT TOP (150) id FROM TestTable ORDER BY NEWID() ) SELECT T1.id, T1.padding FROM TestTable AS T1 WHERE T1.id IN (SELECT id FROM TopKeys) OPTION (MAXDOP 1) ; WITH TestTable AS ( SELECT * FROM dbo.TestTEXT ), TopKeys AS ( SELECT TOP (150) id FROM TestTable ORDER BY NEWID() ) SELECT T1.id, T1.padding FROM TestTable AS T1 WHERE T1.id IN (SELECT id FROM TopKeys) OPTION (MAXDOP 1) ; WITH TestTable AS ( SELECT * FROM dbo.TestMAXOOR ), TopKeys AS ( SELECT TOP (150) id FROM TestTable ORDER BY NEWID() ) SELECT T1.id, T1.padding FROM TestTable AS T1 WHERE T1.id IN (SELECT id FROM TopKeys) OPTION (MAXDOP 1) ; SET STATISTICS IO OFF ; All four queries now return results in much less than a second, with memory grants between 6 and 12MB, and without spilling to tempdb.  The small remaining inefficiency is in reading the id column values from the clustered primary key index.  As a clustered index, it contains all the in-row data at its leaf.  The CHAR and VARCHAR(MAX) tables store the padding column in-row, so id values are separated by a 3999-character column, plus row overhead.  The TEXT and MAXOOR tables store the padding values off-row, so id values in the clustered index leaf are separated by the much-smaller off-row pointer structure.  This difference is reflected in the number of logical page reads performed by the four queries: Table 'TestCHAR' logical reads 25511 lob logical reads 000 Table 'TestMAX'. logical reads 25511 lob logical reads 000 Table 'TestTEXT' logical reads 00412 lob logical reads 597 Table 'TestMAXOOR' logical reads 00413 lob logical reads 446 We can increase the density of the id values by creating a separate nonclustered index on the id column only.  This is the same key as the clustered index, of course, but the nonclustered index will not include the rest of the in-row column data. CREATE UNIQUE NONCLUSTERED INDEX uq1 ON dbo.TestCHAR (id); CREATE UNIQUE NONCLUSTERED INDEX uq1 ON dbo.TestMAX (id); CREATE UNIQUE NONCLUSTERED INDEX uq1 ON dbo.TestTEXT (id); CREATE UNIQUE NONCLUSTERED INDEX uq1 ON dbo.TestMAXOOR (id); The four queries can now use the very dense nonclustered index to quickly scan the id values, sort them by NEWID(), select the 150 ids we want, and then look up the padding data.  The logical reads with the new indexes in place are: Table 'TestCHAR' logical reads 835 lob logical reads 0 Table 'TestMAX' logical reads 835 lob logical reads 0 Table 'TestTEXT' logical reads 686 lob logical reads 597 Table 'TestMAXOOR' logical reads 686 lob logical reads 448 With the new index, all four queries use the same query plan (click to enlarge): Performance Summary: 0.3 seconds elapsed time 6MB memory grant 0MB tempdb usage 1MB sort set 835 logical reads (CHAR, MAX) 686 logical reads (TEXT, MAXOOR) 597 LOB logical reads (TEXT) 448 LOB logical reads (MAXOOR) No sort warning I’ll leave it as an exercise for the reader to work out why trying to eliminate the Key Lookup by adding the padding column to the new nonclustered indexes would be a daft idea Conclusion This post is not about tuning queries that access columns containing big strings.  It isn’t about the internal differences between TEXT and MAX data types either.  It isn’t even about the cool use of UPDATE .WRITE used in the MAXOOR table load.  No, this post is about something else: Many developers might not have tuned our starting example query at all – 5 seconds isn’t that bad, and the original query plan looks reasonable at first glance.  Perhaps the NEWID() function would have been blamed for ‘just being slow’ – who knows.  5 seconds isn’t awful – unless your users expect sub-second responses – but using 250MB of memory and writing 200MB to tempdb certainly is!  If ten sessions ran that query at the same time in production that’s 2.5GB of memory usage and 2GB hitting tempdb.  Of course, not all queries can be rewritten to avoid large memory grants and sort spills using the key-lookup technique in this post, but that’s not the point either. The point of this post is that a basic understanding of execution plans is not enough.  Tuning for logical reads and adding covering indexes is not enough.  If you want to produce high-quality, scalable TSQL that won’t get you paged as soon as it hits production, you need a deep understanding of execution plans, and as much accurate, deep knowledge about SQL Server as you can lay your hands on.  The advanced database developer has a wide range of tools to use in writing queries that perform well in a range of circumstances. By the way, the examples in this post were written for SQL Server 2008.  They will run on 2005 and demonstrate the same principles, but you won’t get the same figures I did because 2005 had a rather nasty bug in the Top N Sort operator.  Fair warning: if you do decide to run the scripts on a 2005 instance (particularly the parallel query) do it before you head out for lunch… This post is dedicated to the people of Christchurch, New Zealand. © 2011 Paul White email: @[email protected] twitter: @SQL_Kiwi

    Read the article

  • Migrating from DD-WRT to Tomato

    - by Collin Allen
    Is it possible to switch to the latest version of Tomato on a router that's already running DD-WRT? Using the default Linksys firmware on my WRT54GL v1.1, I had to upload a micro version of DD-WRT first. I imagine that, since I'm now running third-party firmware, I won't have to do that again to make the switch, but I thought I should check so as not to brick it. This router is taking a back seat to a new AirPort Extreme (for the 'n' capability), but I still want to have the soon-to-be-Tomato device sit between the AirPort Extreme and my modem for the superior traffic graphing.

    Read the article

< Previous Page | 11 12 13 14 15 16 17 18 19 20 21 22  | Next Page >