Search Results

Search found 6028 results on 242 pages for 'total commander'.

Page 33/242 | < Previous Page | 29 30 31 32 33 34 35 36 37 38 39 40  | Next Page >

  • More than 100,000 articles !

    - by developerit
    In one month, we already got more than 100,000, and we continue to crawl! We plan on hitting 250,000 total articles next month. Due to the large amount of data we are gathering, we are planning on updating our SQL stored procedure to improve performance. We may be migrating to SQL Server 2008 Entreprise, as we are currently running on SQL Server 2005 Express Edition… We are at 400 Mb of data, getting more and more close to the 2 Gb limit. Stay tune for more info and browse daily fresh articles about web development.

    Read the article

  • SQL SERVER – Out of the Box – Activty and Performance Reports from SSSMS

    - by pinaldave
    SQL Server management Studio 2008 is wonderful tool and has many different features. Many times, an average user does not use them as they are not aware about these features. Today, we will learn one such feature. SSMS comes with many inbuilt performance and activity reports, but we do not use it to the full potential. Let us see how we can access these standard reports. Connect to SQL Server Node >> Right Click on it >> Go to Reports >> Click on Standard Reports >> Pick Any Report. Click to Enlarge You can see there are many reports, which an average users needs right away, are available there. Let me list all the reports available. Server Dashboard Configuration Changes History Schema Changes History Scheduler Health Memory Consumption Activity – All Blocking Transactions Activity – All Cursors Activity – All Sessions Activity – Top Sessions Activity – Dormant Sessions Activity -  Top Connections Top Transactions by Age Top Transactions by Blocked Transactions Count Top Transactions by Locks Count Performance – Batch Execution Statistics Performance – Object Execution Statistics Performance – Top Queries by Average CPU Time Performance – Top Queries by Average IO Performance – Top Queries by Total CPU Time Performance – Top Queries by Total IO Service Broker Statistics Transactions Log Shipping Status In fact, when you look at the above list, it is fairly clear that they are very thought out and commonly needed reports that are available in SQL Server 2008. Let us run a couple of reports and observe their result. Performance – Top Queries by Total CPU Time Click to Enlarge Memory Consumption Click to Enlarge There are options for custom reports as well, which we can configure. We will learn about them in some other post. Additionally, you can right click on the reports and export in Excel or PDF. I think this tool can really help those who are just looking for some quick details. Does any of you use this feature, or this feature has some limitations and You would like to see more features? Reference : Pinal Dave (http://blog.SQLAuthority.com) Filed under: Pinal Dave, SQL, SQL Authority, SQL Optimization, SQL Performance, SQL Query, SQL Server, SQL Server Management Studio, SQL Tips and Tricks, T SQL, Technology

    Read the article

  • LWJGL - Eclipse error [on hold]

    - by Zarkopafilis
    When I try to run my lwjgl project, an error pops . Here is the log file: # A fatal error has been detected by the Java Runtime Environment: # EXCEPTION_ACCESS_VIOLATION (0xc0000005) at pc=0x6d8fcc0a, pid=5612, tid=900 # JRE version: 6.0_16-b01 Java VM: Java HotSpot(TM) Client VM (14.2-b01 mixed mode windows-x86 ) Problematic frame: V [jvm.dll+0xfcc0a] # If you would like to submit a bug report, please visit: http://java.sun.com/webapps/bugreport/crash.jsp # --------------- T H R E A D --------------- Current thread (0x016b9000): JavaThread "main" [_thread_in_vm, id=900, stack(0x00160000,0x001b0000)] siginfo: ExceptionCode=0xc0000005, reading address 0x00000000 Registers: EAX=0x00000000, EBX=0x00000000, ECX=0x00000006, EDX=0x00000000 ESP=0x001af4d4, EBP=0x001af524, ESI=0x016b9000, EDI=0x016b9110 EIP=0x6d8fcc0a, EFLAGS=0x00010246 Top of Stack: (sp=0x001af4d4) 0x001af4d4: 6da44bd8 016b9110 00000000 001af668 0x001af4e4: ffffffff 22200000 001af620 76ec39c2 0x001af4f4: 001af524 6d801086 0000000b 001afd34 0x001af504: 016b9000 016dd990 016b9000 00000000 0x001af514: 001af5f4 6d9ee000 6d9ef2f0 ffffffff 0x001af524: 001af58c 10008c85 016b9110 00000000 0x001af534: 00000000 000a0554 00000000 00000024 0x001af544: 00000000 00000000 001af6ac 00000000 Instructions: (pc=0x6d8fcc0a) 0x6d8fcbfa: e8 e8 d0 1d 08 00 8b 45 10 c7 45 d8 0b 00 00 00 0x6d8fcc0a: 8b 00 8b 48 08 0f b7 51 26 8b 40 0c 8b 4c 90 20 Stack: [0x00160000,0x001b0000], sp=0x001af4d4, free space=317k Native frames: (J=compiled Java code, j=interpreted, Vv=VM code, C=native code) V [jvm.dll+0xfcc0a] C [lwjgl.dll+0x8c85] C [USER32.dll+0x18876] C [USER32.dll+0x170f4] C [USER32.dll+0x1119e] C [ntdll.dll+0x460ce] C [USER32.dll+0x10e29] C [USER32.dll+0x10e84] C [lwjgl.dll+0x1cf0] j org.lwjgl.opengl.WindowsDisplay.createWindow(Lorg/lwjgl/opengl/DrawableLWJGL;Lorg/lwjgl/opengl/DisplayMode;Ljava/awt/Canvas;II)V+102 j org.lwjgl.opengl.Display.createWindow()V+71 j org.lwjgl.opengl.Display.create(Lorg/lwjgl/opengl/PixelFormat;Lorg/lwjgl/opengl/Drawable;Lorg/lwjgl/opengl/ContextAttribs;)V+72 j org.lwjgl.opengl.Display.create(Lorg/lwjgl/opengl/PixelFormat;)V+12 j org.lwjgl.opengl.Display.create()V+7 j zarkopafilis.koding.io.javafx.Main.main([Ljava/lang/String;)V+16 v ~StubRoutines::call_stub V [jvm.dll+0xecf9c] V [jvm.dll+0x1741e1] V [jvm.dll+0xed01d] V [jvm.dll+0xf5be5] V [jvm.dll+0xfd83d] C [javaw.exe+0x2155] C [javaw.exe+0x833e] C [kernel32.dll+0x51154] C [ntdll.dll+0x5b2b9] C [ntdll.dll+0x5b28c] Java frames: (J=compiled Java code, j=interpreted, Vv=VM code) j org.lwjgl.opengl.WindowsDisplay.nCreateWindow(IIIIZZJ)J+0 j org.lwjgl.opengl.WindowsDisplay.createWindow(Lorg/lwjgl/opengl/DrawableLWJGL;Lorg/lwjgl/opengl/DisplayMode;Ljava/awt/Canvas;II)V+102 j org.lwjgl.opengl.Display.createWindow()V+71 j org.lwjgl.opengl.Display.create(Lorg/lwjgl/opengl/PixelFormat;Lorg/lwjgl/opengl/Drawable;Lorg/lwjgl/opengl/ContextAttribs;)V+72 j org.lwjgl.opengl.Display.create(Lorg/lwjgl/opengl/PixelFormat;)V+12 j org.lwjgl.opengl.Display.create()V+7 j zarkopafilis.koding.io.javafx.Main.main([Ljava/lang/String;)V+16 v ~StubRoutines::call_stub --------------- P R O C E S S --------------- Java Threads: ( = current thread ) 0x0179a400 JavaThread "Low Memory Detector" daemon [_thread_blocked, id=4460, stack(0x0b900000,0x0b950000)] 0x01795400 JavaThread "CompilerThread0" daemon [_thread_blocked, id=5264, stack(0x0b8b0000,0x0b900000)] 0x01790c00 JavaThread "Attach Listener" daemon [_thread_blocked, id=6080, stack(0x0b860000,0x0b8b0000)] 0x01786400 JavaThread "Signal Dispatcher" daemon [_thread_blocked, id=1204, stack(0x0b810000,0x0b860000)] 0x01759c00 JavaThread "Finalizer" daemon [_thread_blocked, id=5772, stack(0x0b7c0000,0x0b810000)] 0x01755000 JavaThread "Reference Handler" daemon [_thread_blocked, id=4696, stack(0x01640000,0x01690000)] =0x016b9000 JavaThread "main" [_thread_in_vm, id=900, stack(0x00160000,0x001b0000)] Other Threads: 0x01751c00 VMThread [stack: 0x015f0000,0x01640000] [id=4052] 0x0179c800 WatcherThread [stack: 0x0b950000,0x0b9a0000] [id=3340] VM state:not at safepoint (normal execution) VM Mutex/Monitor currently owned by a thread: None Heap def new generation total 960K, used 816K [0x037c0000, 0x038c0000, 0x03ca0000) eden space 896K, 91% used [0x037c0000, 0x0388c2c0, 0x038a0000) from space 64K, 0% used [0x038a0000, 0x038a0000, 0x038b0000) to space 64K, 0% used [0x038b0000, 0x038b0000, 0x038c0000) tenured generation total 4096K, used 0K [0x03ca0000, 0x040a0000, 0x077c0000) the space 4096K, 0% used [0x03ca0000, 0x03ca0000, 0x03ca0200, 0x040a0000) compacting perm gen total 12288K, used 2143K [0x077c0000, 0x083c0000, 0x0b7c0000) the space 12288K, 17% used [0x077c0000, 0x079d7e38, 0x079d8000, 0x083c0000) No shared spaces configured. Dynamic libraries: 0x00400000 - 0x00424000 C:\Program Files\Java\jre6\bin\javaw.exe 0x77550000 - 0x7768e000 C:\Windows\SYSTEM32\ntdll.dll 0x75a80000 - 0x75b54000 C:\Windows\system32\kernel32.dll 0x758d0000 - 0x7591b000 C:\Windows\system32\KERNELBASE.dll 0x759e0000 - 0x75a80000 C:\Windows\system32\ADVAPI32.dll 0x76070000 - 0x7611c000 C:\Windows\system32\msvcrt.dll 0x77250000 - 0x77269000 C:\Windows\SYSTEM32\sechost.dll 0x771a0000 - 0x77241000 C:\Windows\system32\RPCRT4.dll 0x76eb0000 - 0x76f79000 C:\Windows\system32\USER32.dll 0x76e60000 - 0x76eae000 C:\Windows\system32\GDI32.dll 0x77770000 - 0x7777a000 C:\Windows\system32\LPK.dll 0x75fd0000 - 0x7606e000 C:\Windows\system32\USP10.dll 0x770b0000 - 0x770cf000 C:\Windows\system32\IMM32.DLL 0x770d0000 - 0x7719c000 C:\Windows\system32\MSCTF.dll 0x7c340000 - 0x7c396000 C:\Program Files\Java\jre6\bin\msvcr71.dll 0x6d800000 - 0x6da8b000 C:\Program Files\Java\jre6\bin\client\jvm.dll 0x73a00000 - 0x73a32000 C:\Windows\system32\WINMM.dll 0x75610000 - 0x7565b000 C:\Windows\system32\apphelp.dll 0x6d7b0000 - 0x6d7bc000 C:\Program Files\Java\jre6\bin\verify.dll 0x6d330000 - 0x6d34f000 C:\Program Files\Java\jre6\bin\java.dll 0x6d290000 - 0x6d298000 C:\Program Files\Java\jre6\bin\hpi.dll 0x776e0000 - 0x776e5000 C:\Windows\system32\PSAPI.DLL 0x6d7f0000 - 0x6d7ff000 C:\Program Files\Java\jre6\bin\zip.dll 0x10000000 - 0x1004c000 C:\Users\theo\Desktop\workspace\JavaFX1\lib\natives\windows\lwjgl.dll 0x5d170000 - 0x5d238000 C:\Windows\system32\OPENGL32.dll 0x6e7b0000 - 0x6e7d2000 C:\Windows\system32\GLU32.dll 0x70620000 - 0x70707000 C:\Windows\system32\DDRAW.dll 0x70610000 - 0x70616000 C:\Windows\system32\DCIMAN32.dll 0x75b60000 - 0x75cfd000 C:\Windows\system32\SETUPAPI.dll 0x759b0000 - 0x759d7000 C:\Windows\system32\CFGMGR32.dll 0x76d70000 - 0x76dff000 C:\Windows\system32\OLEAUT32.dll 0x75db0000 - 0x75f0c000 C:\Windows\system32\ole32.dll 0x758b0000 - 0x758c2000 C:\Windows\system32\DEVOBJ.dll 0x74060000 - 0x74073000 C:\Windows\system32\dwmapi.dll 0x74b60000 - 0x74b69000 C:\Windows\system32\VERSION.dll 0x745f0000 - 0x7478e000 C:\Windows\WinSxS\x86_microsoft.windows.common-controls_6595b64144ccf1df_6.0.7600.16661_none_420fe3fa2b8113bd\COMCTL32.dll 0x75d50000 - 0x75da7000 C:\Windows\system32\SHLWAPI.dll 0x74370000 - 0x743b0000 C:\Windows\system32\uxtheme.dll 0x22200000 - 0x22206000 C:\Program Files\ESET\ESET Smart Security\eplgHooks.dll VM Arguments: jvm_args: -Djava.library.path=C:\Users\theo\Desktop\workspace\JavaFX1\lib\natives\windows -Dfile.encoding=Cp1253 java_command: zarkopafilis.koding.io.javafx.Main Launcher Type: SUN_STANDARD Environment Variables: PATH=C:/Program Files/Java/jre6/bin/client;C:/Program Files/Java/jre6/bin;C:/Program Files/Java/jre6/lib/i386;C:\Perl\site\bin;C:\Perl\bin;C:\Ruby200\bin;C:\Program Files\Common Files\Microsoft Shared\Windows Live;C:\Windows\system32;C:\Windows;C:\Windows\System32\Wbem;C:\Windows\System32\WindowsPowerShell\v1.0\;C:\Program Files\Windows Live\Shared;C:\Users\theo\Desktop\eclipse; USERNAME=theo OS=Windows_NT PROCESSOR_IDENTIFIER=x86 Family 6 Model 37 Stepping 5, GenuineIntel --------------- S Y S T E M --------------- OS: Windows 7 Build 7600 CPU:total 4 (8 cores per cpu, 2 threads per core) family 6 model 37 stepping 5, cmov, cx8, fxsr, mmx, sse, sse2, sse3, ssse3, sse4.1, sse4.2, ht Memory: 4k page, physical 2097151k(1257972k free), swap 4194303k(4194303k free) vm_info: Java HotSpot(TM) Client VM (14.2-b01) for windows-x86 JRE (1.6.0_16-b01), built on Jul 31 2009 11:26:58 by "java_re" with MS VC++ 7.1 time: Wed Oct 23 22:00:12 2013 elapsed time: 0 seconds Code: Display.setDisplayMode(new DisplayMode(800,600)); Display.create();//Error here I am using JDK 6

    Read the article

  • Installation doesn't detect existing partitions

    - by retrac1324
    I am trying to install Ubuntu 11.10 in a dual boot with my existing Windows 7 but the installer does not detect any existing partitions. I have tried resetting my BCD using EasyBCD and doing fixmbr from the Windows startup disc. A while ago I had to use TestDisk to recover my partition table so this might be the cause but I have installed Ubuntu and Windows many times before with no problems. fdisk -l output: Disk /dev/sda: 640.1 GB, 640135028736 bytes 255 heads, 63 sectors/track, 77825 cylinders, total 1250263728 sectors Units = sectors of 1 * 512 = 512 bytes Sector size (logical/physical): 512 bytes / 512 bytes I/O size (minimum/optimal): 512 bytes / 512 bytes Disk identifier: 0x360555e5 Device Boot Start End Blocks Id System /dev/sda1 * 2048 1250274689 625136321 7 HPFS/NTFS/exFAT Disk /dev/sdf: 7803 MB, 7803174912 bytes 255 heads, 63 sectors/track, 948 cylinders, total 15240576 sectors Units = sectors of 1 * 512 = 512 bytes Sector size (logical/physical): 512 bytes / 512 bytes I/O size (minimum/optimal): 512 bytes / 512 bytes Disk identifier: 0x6f795a8d Device Boot Start End Blocks Id System /dev/sdf1 * 63 15240575 7620256+ c W95 FAT32 (LBA)

    Read the article

  • Parallelism in .NET – Part 11, Divide and Conquer via Parallel.Invoke

    - by Reed
    Many algorithms are easily written to work via recursion.  For example, most data-oriented tasks where a tree of data must be processed are much more easily handled by starting at the root, and recursively “walking” the tree.  Some algorithms work this way on flat data structures, such as arrays, as well.  This is a form of divide and conquer: an algorithm design which is based around breaking up a set of work recursively, “dividing” the total work in each recursive step, and “conquering” the work when the remaining work is small enough to be solved easily. Recursive algorithms, especially ones based on a form of divide and conquer, are often a very good candidate for parallelization. This is apparent from a common sense standpoint.  Since we’re dividing up the total work in the algorithm, we have an obvious, built-in partitioning scheme.  Once partitioned, the data can be worked upon independently, so there is good, clean isolation of data. Implementing this type of algorithm is fairly simple.  The Parallel class in .NET 4 includes a method suited for this type of operation: Parallel.Invoke.  This method works by taking any number of delegates defined as an Action, and operating them all in parallel.  The method returns when every delegate has completed: Parallel.Invoke( () => { Console.WriteLine("Action 1 executing in thread {0}", Thread.CurrentThread.ManagedThreadId); }, () => { Console.WriteLine("Action 2 executing in thread {0}", Thread.CurrentThread.ManagedThreadId); }, () => { Console.WriteLine("Action 3 executing in thread {0}", Thread.CurrentThread.ManagedThreadId); } ); .csharpcode, .csharpcode pre { font-size: small; color: black; font-family: consolas, "Courier New", courier, monospace; background-color: #ffffff; /*white-space: pre;*/ } .csharpcode pre { margin: 0em; } .csharpcode .rem { color: #008000; } .csharpcode .kwrd { color: #0000ff; } .csharpcode .str { color: #006080; } .csharpcode .op { color: #0000c0; } .csharpcode .preproc { color: #cc6633; } .csharpcode .asp { background-color: #ffff00; } .csharpcode .html { color: #800000; } .csharpcode .attr { color: #ff0000; } .csharpcode .alt { background-color: #f4f4f4; width: 100%; margin: 0em; } .csharpcode .lnum { color: #606060; } Running this simple example demonstrates the ease of using this method.  For example, on my system, I get three separate thread IDs when running the above code.  By allowing any number of delegates to be executed directly, concurrently, the Parallel.Invoke method provides us an easy way to parallelize any algorithm based on divide and conquer.  We can divide our work in each step, and execute each task in parallel, recursively. For example, suppose we wanted to implement our own quicksort routine.  The quicksort algorithm can be designed based on divide and conquer.  In each iteration, we pick a pivot point, and use that to partition the total array.  We swap the elements around the pivot, then recursively sort the lists on each side of the pivot.  For example, let’s look at this simple, sequential implementation of quicksort: public static void QuickSort<T>(T[] array) where T : IComparable<T> { QuickSortInternal(array, 0, array.Length - 1); } private static void QuickSortInternal<T>(T[] array, int left, int right) where T : IComparable<T> { if (left >= right) { return; } SwapElements(array, left, (left + right) / 2); int last = left; for (int current = left + 1; current <= right; ++current) { if (array[current].CompareTo(array[left]) < 0) { ++last; SwapElements(array, last, current); } } SwapElements(array, left, last); QuickSortInternal(array, left, last - 1); QuickSortInternal(array, last + 1, right); } static void SwapElements<T>(T[] array, int i, int j) { T temp = array[i]; array[i] = array[j]; array[j] = temp; } Here, we implement the quicksort algorithm in a very common, divide and conquer approach.  Running this against the built-in Array.Sort routine shows that we get the exact same answers (although the framework’s sort routine is slightly faster).  On my system, for example, I can use framework’s sort to sort ten million random doubles in about 7.3s, and this implementation takes about 9.3s on average. Looking at this routine, though, there is a clear opportunity to parallelize.  At the end of QuickSortInternal, we recursively call into QuickSortInternal with each partition of the array after the pivot is chosen.  This can be rewritten to use Parallel.Invoke by simply changing it to: // Code above is unchanged... SwapElements(array, left, last); Parallel.Invoke( () => QuickSortInternal(array, left, last - 1), () => QuickSortInternal(array, last + 1, right) ); } This routine will now run in parallel.  When executing, we now see the CPU usage across all cores spike while it executes.  However, there is a significant problem here – by parallelizing this routine, we took it from an execution time of 9.3s to an execution time of approximately 14 seconds!  We’re using more resources as seen in the CPU usage, but the overall result is a dramatic slowdown in overall processing time. This occurs because parallelization adds overhead.  Each time we split this array, we spawn two new tasks to parallelize this algorithm!  This is far, far too many tasks for our cores to operate upon at a single time.  In effect, we’re “over-parallelizing” this routine.  This is a common problem when working with divide and conquer algorithms, and leads to an important observation: When parallelizing a recursive routine, take special care not to add more tasks than necessary to fully utilize your system. This can be done with a few different approaches, in this case.  Typically, the way to handle this is to stop parallelizing the routine at a certain point, and revert back to the serial approach.  Since the first few recursions will all still be parallelized, our “deeper” recursive tasks will be running in parallel, and can take full advantage of the machine.  This also dramatically reduces the overhead added by parallelizing, since we’re only adding overhead for the first few recursive calls.  There are two basic approaches we can take here.  The first approach would be to look at the total work size, and if it’s smaller than a specific threshold, revert to our serial implementation.  In this case, we could just check right-left, and if it’s under a threshold, call the methods directly instead of using Parallel.Invoke. The second approach is to track how “deep” in the “tree” we are currently at, and if we are below some number of levels, stop parallelizing.  This approach is a more general-purpose approach, since it works on routines which parse trees as well as routines working off of a single array, but may not work as well if a poor partitioning strategy is chosen or the tree is not balanced evenly. This can be written very easily.  If we pass a maxDepth parameter into our internal routine, we can restrict the amount of times we parallelize by changing the recursive call to: // Code above is unchanged... SwapElements(array, left, last); if (maxDepth < 1) { QuickSortInternal(array, left, last - 1, maxDepth); QuickSortInternal(array, last + 1, right, maxDepth); } else { --maxDepth; Parallel.Invoke( () => QuickSortInternal(array, left, last - 1, maxDepth), () => QuickSortInternal(array, last + 1, right, maxDepth)); } We no longer allow this to parallelize indefinitely – only to a specific depth, at which time we revert to a serial implementation.  By starting the routine with a maxDepth equal to Environment.ProcessorCount, we can restrict the total amount of parallel operations significantly, but still provide adequate work for each processing core. With this final change, my timings are much better.  On average, I get the following timings: Framework via Array.Sort: 7.3 seconds Serial Quicksort Implementation: 9.3 seconds Naive Parallel Implementation: 14 seconds Parallel Implementation Restricting Depth: 4.7 seconds Finally, we are now faster than the framework’s Array.Sort implementation.

    Read the article

  • Loading PNGs into OpenGL performance issues - Java & JOGL much slower than C# & Tao.OpenGL

    - by Edward Cresswell
    I am noticing a large performance difference between Java & JOGL and C# & Tao.OpenGL when both loading PNGs from storage into memory, and when loading that BufferedImage (java) or Bitmap (C# - both are PNGs on hard drive) 'into' OpenGL. This difference is quite large, so I assumed I was doing something wrong, however after quite a lot of searching and trying different loading techniques I've been unable to reduce this difference. With Java I get an image loaded in 248ms and loaded into OpenGL in 728ms The same on C# takes 54ms to load the image, and 34ms to load/create texture. The image in question above is a PNG containing transparency, sized 7200x255, used for a 2D animated sprite. I realise the size is really quite ridiculous and am considering cutting up the sprite, however the large difference is still there (and confusing). On the Java side the code looks like this: BufferedImage image = ImageIO.read(new File(fileName)); texture = TextureIO.newTexture(image, false); texture.setTexParameteri(GL.GL_TEXTURE_MIN_FILTER, GL.GL_LINEAR); texture.setTexParameteri(GL.GL_TEXTURE_MAG_FILTER, GL.GL_LINEAR); The C# code uses: Bitmap t = new Bitmap(fileName); t.RotateFlip(RotateFlipType.RotateNoneFlipY); Rectangle r = new Rectangle(0, 0, t.Width, t.Height); BitmapData bd = t.LockBits(r, ImageLockMode.ReadOnly, PixelFormat.Format32bppArgb); Gl.glBindTexture(Gl.GL_TEXTURE_2D, tID); Gl.glTexImage2D(Gl.GL_TEXTURE_2D, 0, Gl.GL_RGBA, t.Width, t.Height, 0, Gl.GL_BGRA, Gl.GL_UNSIGNED_BYTE, bd.Scan0); Gl.glTexParameteri(Gl.GL_TEXTURE_2D, Gl.GL_TEXTURE_MIN_FILTER, Gl.GL_LINEAR); Gl.glTexParameteri(Gl.GL_TEXTURE_2D, Gl.GL_TEXTURE_MAG_FILTER, Gl.GL_LINEAR); t.UnlockBits(bd); t.Dispose(); After quite a lot of testing I can only come to the conclusion that Java/JOGL is just slower here - PNG reading might not be as quick, or that I'm still doing something wrong. Thanks. Edit2: I have found that creating a new BufferedImage with format TYPE_INT_ARGB_PRE decreases OpenGL texture load time by almost half - this includes having to create the new BufferedImage, getting the Graphics2D from it and then rendering the previously loaded image to it. Edit3: Benchmark results for 5 variations. I wrote a small benchmarking tool, the following results come from loading a set of 33 pngs, most are very wide, 5 times. testStart: ImageIO.read(file) -> TextureIO.newTexture(image) result: avg = 10250ms, total = 51251 testStart: ImageIO.read(bis) -> TextureIO.newTexture(image) result: avg = 10029ms, total = 50147 testStart: ImageIO.read(file) -> TextureIO.newTexture(argbImage) result: avg = 5343ms, total = 26717 testStart: ImageIO.read(bis) -> TextureIO.newTexture(argbImage) result: avg = 5534ms, total = 27673 testStart: TextureIO.newTexture(file) result: avg = 10395ms, total = 51979 ImageIO.read(bis) refers to the technique described in James Branigan's answer below. argbImage refers to the technique described in my previous edit: img = ImageIO.read(file); argbImg = new BufferedImage(img.getWidth(), img.getHeight(), TYPE_INT_ARGB_PRE); g = argbImg.createGraphics(); g.drawImage(img, 0, 0, null); texture = TextureIO.newTexture(argbImg, false); Any more methods of loading (either images from file, or images to OpenGL) would be appreciated, I will update these benchmarks.

    Read the article

  • Option Trading: Getting the most out of the event session options

    - by extended_events
    You can control different aspects of how an event session behaves by setting the event session options as part of the CREATE EVENT SESSION DDL. The default settings for the event session options are designed to handle most of the common event collection situations so I generally recommend that you just use the defaults. Like everything in the real world though, there are going to be a handful of “special cases” that require something different. This post focuses on identifying the special cases and the correct use of the options to accommodate those cases. There is a reason it’s called Default The default session options specify a total event buffer size of 4 MB with a 30 second latency. Translating this into human terms; this means that our default behavior is that the system will start processing events from the event buffer when we reach about 1.3 MB of events or after 30 seconds, which ever comes first. Aside: What’s up with the 1.3 MB, I thought you said the buffer was 4 MB?The Extended Events engine takes the total buffer size specified by MAX_MEMORY (4MB by default) and divides it into 3 equally sized buffers. This is done so that a session can be publishing events to one buffer while other buffers are being processed. There are always at least three buffers; how to get more than three is covered later. Using this configuration, the Extended Events engine can “keep up” with most event sessions on standard workloads. Why is this? The fact is that most events are small, really small; on the order of a couple hundred bytes. Even when you start considering events that carry dynamically sized data (eg. binary, text, etc.) or adding actions that collect additional data, the total size of the event is still likely to be pretty small. This means that each buffer can likely hold thousands of events before it has to be processed. When the event buffers are finally processed there is an economy of scale achieved since most targets support bulk processing of the events so they are processed at the buffer level rather than the individual event level. When all this is working together it’s more likely that a full buffer will be processed and put back into the ready queue before the remaining buffers (remember, there are at least three) are full. I know what you’re going to say: “My server is exceptional! My workload is so massive it defies categorization!” OK, maybe you weren’t going to say that exactly, but you were probably thinking it. The point is that there are situations that won’t be covered by the Default, but that’s a good place to start and this post assumes you’ve started there so that you have something to look at in order to determine if you do have a special case that needs different settings. So let’s get to the special cases… What event just fired?! How about now?! Now?! If you believe the commercial adage from Heinz Ketchup (Heinz Slow Good Ketchup ad on You Tube), some things are worth the wait. This is not a belief held by most DBAs, particularly DBAs who are looking for an answer to a troubleshooting question fast. If you’re one of these anxious DBAs, or maybe just a Program Manager doing a demo, then 30 seconds might be longer than you’re comfortable waiting. If you find yourself in this situation then consider changing the MAX_DISPATCH_LATENCY option for your event session. This option will force the event buffers to be processed based on your time schedule. This option only makes sense for the asynchronous targets since those are the ones where we allow events to build up in the event buffer – if you’re using one of the synchronous targets this option isn’t relevant. Avoid forgotten events by increasing your memory Have you ever had one of those days where you keep forgetting things? That can happen in Extended Events too; we call it dropped events. In order to optimizes for server performance and help ensure that the Extended Events doesn’t block the server if to drop events that can’t be published to a buffer because the buffer is full. You can determine if events are being dropped from a session by querying the dm_xe_sessions DMV and looking at the dropped_event_count field. Aside: Should you care if you’re dropping events?Maybe not – think about why you’re collecting data in the first place and whether you’re really going to miss a few dropped events. For example, if you’re collecting query duration stats over thousands of executions of a query it won’t make a huge difference to miss a couple executions. Use your best judgment. If you find that your session is dropping events it means that the event buffer is not large enough to handle the volume of events that are being published. There are two ways to address this problem. First, you could collect fewer events – examine you session to see if you are over collecting. Do you need all the actions you’ve specified? Could you apply a predicate to be more specific about when you fire the event? Assuming the session is defined correctly, the next option is to change the MAX_MEMORY option to a larger number. Picking the right event buffer size might take some trial and error, but a good place to start is with the number of dropped events compared to the number you’ve collected. Aside: There are three different behaviors for dropping events that you specify using the EVENT_RETENTION_MODE option. The default is to allow single event loss and you should stick with this setting since it is the best choice for keeping the impact on server performance low.You’ll be tempted to use the setting to not lose any events (NO_EVENT_LOSS) – resist this urge since it can result in blocking on the server. If you’re worried that you’re losing events you should be increasing your event buffer memory as described in this section. Some events are too big to fail A less common reason for dropping an event is when an event is so large that it can’t fit into the event buffer. Even though most events are going to be small, you might find a condition that occasionally generates a very large event. You can determine if your session is dropping large events by looking at the dm_xe_sessions DMV once again, this time check the largest_event_dropped_size. If this value is larger than the size of your event buffer [remember, the size of your event buffer, by default, is max_memory / 3] then you need a large event buffer. To specify a large event buffer you set the MAX_EVENT_SIZE option to a value large enough to fit the largest event dropped based on data from the DMV. When you set this option the Extended Events engine will create two buffers of this size to accommodate these large events. As an added bonus (no extra charge) the large event buffer will also be used to store normal events in the cases where the normal event buffers are all full and waiting to be processed. (Note: This is just a side-effect, not the intended use. If you’re dropping many normal events then you should increase your normal event buffer size.) Partitioning: moving your events to a sub-division Earlier I alluded to the fact that you can configure your event session to use more than the standard three event buffers – this is called partitioning and is controlled by the MEMORY_PARTITION_MODE option. The result of setting this option is fairly easy to explain, but knowing when to use it is a bit more art than science. First the science… You can configure partitioning in three ways: None, Per NUMA Node & Per CPU. This specifies the location where sets of event buffers are created with fairly obvious implication. There are rules we follow for sub-dividing the total memory (specified by MAX_MEMORY) between all the event buffers that are specific to the mode used: None: 3 buffers (fixed)Node: 3 * number_of_nodesCPU: 2.5 * number_of_cpus Here are some examples of what this means for different Node/CPU counts: Configuration None Node CPU 2 CPUs, 1 Node 3 buffers 3 buffers 5 buffers 6 CPUs, 2 Node 3 buffers 6 buffers 15 buffers 40 CPUs, 5 Nodes 3 buffers 15 buffers 100 buffers   Aside: Buffer size on multi-processor computersAs the number of Nodes or CPUs increases, the size of the event buffer gets smaller because the total memory is sub-divided into more pieces. The defaults will hold up to this for a while since each buffer set is holding events only from the Node or CPU that it is associated with, but at some point the buffers will get too small and you’ll either see events being dropped or you’ll get an error when you create your session because you’re below the minimum buffer size. Increase the MAX_MEMORY setting to an appropriate number for the configuration. The most likely reason to start partitioning is going to be related to performance. If you notice that running an event session is impacting the performance of your server beyond a reasonably expected level [Yes, there is a reasonably expected level of work required to collect events.] then partitioning might be an answer. Before you partition you might want to check a few other things: Is your event retention set to NO_EVENT_LOSS and causing blocking? (I told you not to do this.) Consider changing your event loss mode or increasing memory. Are you over collecting and causing more work than necessary? Consider adding predicates to events or removing unnecessary events and actions from your session. Are you writing the file target to the same slow disk that you use for TempDB and your other high activity databases? <kidding> <not really> It’s always worth considering the end to end picture – if you’re writing events to a file you can be impacted by I/O, network; all the usual stuff. Assuming you’ve ruled out the obvious (and not so obvious) issues, there are performance conditions that will be addressed by partitioning. For example, it’s possible to have a successful event session (eg. no dropped events) but still see a performance impact because you have many CPUs all attempting to write to the same free buffer and having to wait in line to finish their work. This is a case where partitioning would relieve the contention between the different CPUs and likely reduce the performance impact cause by the event session. There is no DMV you can check to find these conditions – sorry – that’s where the art comes in. This is  largely a matter of experimentation. On the bright side you probably won’t need to to worry about this level of detail all that often. The performance impact of Extended Events is significantly lower than what you may be used to with SQL Trace. You will likely only care about the impact if you are trying to set up a long running event session that will be part of your everyday workload – sessions used for short term troubleshooting will likely fall into the “reasonably expected impact” category. Hey buddy – I think you forgot something OK, there are two options I didn’t cover: STARTUP_STATE & TRACK_CAUSALITY. If you want your event sessions to start automatically when the server starts, set the STARTUP_STATE option to ON. (Now there is only one option I didn’t cover.) I’m going to leave causality for another post since it’s not really related to session behavior, it’s more about event analysis. - Mike Share this post: email it! | bookmark it! | digg it! | reddit! | kick it! | live it!

    Read the article

  • PHP pages working slow from time to time

    - by user1038179
    I have VPS with limit of 2GB of ram and 8 CPU cores. I have 5 sites on that VPS (one of them is just for testing, no visitors exept me). All 5 sites are image galleries, like wallpaper sites. Last week I noticed problem on one site (main domain, used for name servers, and also with most traffic, visitors). That site has two image galleries, one is old static html gallery made few years ago and another, main, is powered by ZENPhoto CMS. Also I have that same gallery CMS on another two sites on that same VPS (on one running site and on one just for testing site). On other two sites I have diferent PHP driven gallery. Problem is that after some time (it vary from 10 minutes to few hours after apache restart), loading of pages on main site becomes very slow, or I get 503 Service Temporarily Unavailable error. So pages becomes unavailable. But just that part with new CMS gallery, old part of site with static html pages are working fast and just fine. Also other two sites with same CMS gallery and other two with different PHP driven gallery are working fine and fast at the same time. I thought it must be something with CMS on that main site, because other sites are working nice. Then I tryed to open contact and guest book pages on that main site which are outside of that CMS but also PHP pages, and they do not load too, but that same contact php scipts are working on other sites at the same time. So, when site starts to hangs, ONLY PHP generated content is not working, like I said other static pages are working. And, ONLY on that one main site I have problems. Then I need to restart Apache, after restart everything is vorking nice and fast, for some time, than again, just PHP pages on main site are becomming slower. If I do not restart apache that slowness take some time (several minutes, hours, depending ot traffic) and during that time PHP diven content is loading very slow or unavailable on that site. After sime time, on moments everything start to work and is fast again for some time, and again. In hours with more traffic PHP content is loading slowly or it is unavailable, in hours with less traffic it is sometimes fast and sometimes little bit slower than usually. And ones again, only on that main site, and only PHP driven pages, static pages are working fast even in most traffic hours also other sites with even same CMS are working fast. Currently I have about 7000 unique visitors on that site but site worked nice even with 11500 visitors per day. And about 17000 in total visitors on VPS, all sites ( about 3 pages per unique visitor). When site start to slow down sometimes in apache status I can see something like this: mod_fcgid status: Total FastCGI processes: 37 Process: php5 (/usr/local/cpanel/cgi-sys/php5)Pid Active Idle Accesses State 11300 39 28 7 Working 11274 47 28 7 Working 11296 40 29 3 Working 11283 45 30 3 Working 11304 36 31 1 Working 11282 46 32 3 Working 11292 42 33 1 Working 11289 44 34 1 Working 11305 35 35 0 Working 11273 48 36 2 Working 11280 47 39 1 Working 10125 133 40 12 Exiting(communication error) 11294 41 41 1 Exiting(communication error) 11277 47 42 2 Exiting(communication error) 11291 43 43 1 Exiting(communication error) 10187 108 43 10 Exiting(communication error) 10209 95 44 7 Exiting(communication error) 10171 113 44 5 Exiting(communication error) 11275 47 47 1 Exiting(communication error) 10144 125 48 8 Exiting(communication error) 10086 149 48 20 Exiting(communication error) 10212 94 49 5 Exiting(communication error) 10158 118 49 5 Exiting(communication error) 10169 114 50 4 Exiting(communication error) 10105 141 50 16 Exiting(communication error) 10094 146 50 15 Exiting(communication error) 10115 139 51 17 Exiting(communication error) 10213 93 51 9 Exiting(communication error) 10197 103 51 7 Exiting(communication error) Process: php5 (/usr/local/cpanel/cgi-sys/php5)Pid Active Idle Accesses State 7983 1079 2 149 Ready 7979 1079 11 151 Ready Process: php5 (/usr/local/cpanel/cgi-sys/php5)Pid Active Idle Accesses State 7990 1066 0 57 Ready 8001 1031 64 35 Ready 7999 1032 94 29 Ready 8000 1031 91 36 Ready 8002 1029 34 52 Ready Process: php5 (/usr/local/cpanel/cgi-sys/php5)Pid Active Idle Accesses State 7991 1064 29 115 Ready When it is working nicly there is no lines with "Exiting(communication error)" Active and Idle are time active and time since last request, in seconds. Here are system info. Sysem info: Total processors: 8 Processor #1 Vendor GenuineIntel Name Intel(R) Xeon(R) CPU E5440 @ 2.83GHz Speed 88.320 MHz Cache 6144 KB All other seven are the same. System Information Linux vps.nnnnnnnnnnnnnnnnn.nnn 2.6.18-028stab099.3 #1 SMP Wed Mar 7 15:20:22 MSK 2012 x86_64 x86_64 x86_64 GNU/Linux Current Memory Usage total used free shared buffers cached Mem: 8388608 882164 7506444 0 0 0 -/+ buffers/cache: 882164 7506444 Swap: 0 0 0 Total: 8388608 882164 7506444 Current Disk Usage Filesystem Size Used Avail Use% Mounted on /dev/vzfs 100G 34G 67G 34% / none System Details: Running on: Apache/2.2.22 System info: (Unix) mod_ssl/2.2.22 OpenSSL/0.9.8e-fips-rhel5 DAV/2 mod_auth_passthrough/2.1 mod_bwlimited/1.4 FrontPage/5.0.2.2635 mod_fcgid/2.3.6 Powered by: PHP/5.3.10 Current Configuration Default PHP Version (.php files) 5 PHP 5 Handler fcgi PHP 4 Handler suphp Apache suEXEC on Apache Ruid2 off PHP 4 Handler suphp Apache suEXEC on Apache Configuration The following settings have been saved: fileetag: All keepalive: On keepalivetimeout: 3 maxclients: 150 maxkeepaliverequests: 10 maxrequestsperchild: 10000 maxspareservers: 10 minspareservers: 5 root_options: ExecCGI, FollowSymLinks, Includes, IncludesNOEXEC, Indexes, MultiViews, SymLinksIfOwnerMatch serverlimit: 256 serversignature: Off servertokens: Full sslciphersuite: ALL:!ADH:RC4+RSA:+HIGH:+MEDIUM:-LOW:-SSLv2:-EXP:!kEDH startservers: 5 timeout: 30 I hope, I explained my problem nicely. Any help would be nice.

    Read the article

  • Option Trading: Getting the most out of the event session options

    - by extended_events
    You can control different aspects of how an event session behaves by setting the event session options as part of the CREATE EVENT SESSION DDL. The default settings for the event session options are designed to handle most of the common event collection situations so I generally recommend that you just use the defaults. Like everything in the real world though, there are going to be a handful of “special cases” that require something different. This post focuses on identifying the special cases and the correct use of the options to accommodate those cases. There is a reason it’s called Default The default session options specify a total event buffer size of 4 MB with a 30 second latency. Translating this into human terms; this means that our default behavior is that the system will start processing events from the event buffer when we reach about 1.3 MB of events or after 30 seconds, which ever comes first. Aside: What’s up with the 1.3 MB, I thought you said the buffer was 4 MB?The Extended Events engine takes the total buffer size specified by MAX_MEMORY (4MB by default) and divides it into 3 equally sized buffers. This is done so that a session can be publishing events to one buffer while other buffers are being processed. There are always at least three buffers; how to get more than three is covered later. Using this configuration, the Extended Events engine can “keep up” with most event sessions on standard workloads. Why is this? The fact is that most events are small, really small; on the order of a couple hundred bytes. Even when you start considering events that carry dynamically sized data (eg. binary, text, etc.) or adding actions that collect additional data, the total size of the event is still likely to be pretty small. This means that each buffer can likely hold thousands of events before it has to be processed. When the event buffers are finally processed there is an economy of scale achieved since most targets support bulk processing of the events so they are processed at the buffer level rather than the individual event level. When all this is working together it’s more likely that a full buffer will be processed and put back into the ready queue before the remaining buffers (remember, there are at least three) are full. I know what you’re going to say: “My server is exceptional! My workload is so massive it defies categorization!” OK, maybe you weren’t going to say that exactly, but you were probably thinking it. The point is that there are situations that won’t be covered by the Default, but that’s a good place to start and this post assumes you’ve started there so that you have something to look at in order to determine if you do have a special case that needs different settings. So let’s get to the special cases… What event just fired?! How about now?! Now?! If you believe the commercial adage from Heinz Ketchup (Heinz Slow Good Ketchup ad on You Tube), some things are worth the wait. This is not a belief held by most DBAs, particularly DBAs who are looking for an answer to a troubleshooting question fast. If you’re one of these anxious DBAs, or maybe just a Program Manager doing a demo, then 30 seconds might be longer than you’re comfortable waiting. If you find yourself in this situation then consider changing the MAX_DISPATCH_LATENCY option for your event session. This option will force the event buffers to be processed based on your time schedule. This option only makes sense for the asynchronous targets since those are the ones where we allow events to build up in the event buffer – if you’re using one of the synchronous targets this option isn’t relevant. Avoid forgotten events by increasing your memory Have you ever had one of those days where you keep forgetting things? That can happen in Extended Events too; we call it dropped events. In order to optimizes for server performance and help ensure that the Extended Events doesn’t block the server if to drop events that can’t be published to a buffer because the buffer is full. You can determine if events are being dropped from a session by querying the dm_xe_sessions DMV and looking at the dropped_event_count field. Aside: Should you care if you’re dropping events?Maybe not – think about why you’re collecting data in the first place and whether you’re really going to miss a few dropped events. For example, if you’re collecting query duration stats over thousands of executions of a query it won’t make a huge difference to miss a couple executions. Use your best judgment. If you find that your session is dropping events it means that the event buffer is not large enough to handle the volume of events that are being published. There are two ways to address this problem. First, you could collect fewer events – examine you session to see if you are over collecting. Do you need all the actions you’ve specified? Could you apply a predicate to be more specific about when you fire the event? Assuming the session is defined correctly, the next option is to change the MAX_MEMORY option to a larger number. Picking the right event buffer size might take some trial and error, but a good place to start is with the number of dropped events compared to the number you’ve collected. Aside: There are three different behaviors for dropping events that you specify using the EVENT_RETENTION_MODE option. The default is to allow single event loss and you should stick with this setting since it is the best choice for keeping the impact on server performance low.You’ll be tempted to use the setting to not lose any events (NO_EVENT_LOSS) – resist this urge since it can result in blocking on the server. If you’re worried that you’re losing events you should be increasing your event buffer memory as described in this section. Some events are too big to fail A less common reason for dropping an event is when an event is so large that it can’t fit into the event buffer. Even though most events are going to be small, you might find a condition that occasionally generates a very large event. You can determine if your session is dropping large events by looking at the dm_xe_sessions DMV once again, this time check the largest_event_dropped_size. If this value is larger than the size of your event buffer [remember, the size of your event buffer, by default, is max_memory / 3] then you need a large event buffer. To specify a large event buffer you set the MAX_EVENT_SIZE option to a value large enough to fit the largest event dropped based on data from the DMV. When you set this option the Extended Events engine will create two buffers of this size to accommodate these large events. As an added bonus (no extra charge) the large event buffer will also be used to store normal events in the cases where the normal event buffers are all full and waiting to be processed. (Note: This is just a side-effect, not the intended use. If you’re dropping many normal events then you should increase your normal event buffer size.) Partitioning: moving your events to a sub-division Earlier I alluded to the fact that you can configure your event session to use more than the standard three event buffers – this is called partitioning and is controlled by the MEMORY_PARTITION_MODE option. The result of setting this option is fairly easy to explain, but knowing when to use it is a bit more art than science. First the science… You can configure partitioning in three ways: None, Per NUMA Node & Per CPU. This specifies the location where sets of event buffers are created with fairly obvious implication. There are rules we follow for sub-dividing the total memory (specified by MAX_MEMORY) between all the event buffers that are specific to the mode used: None: 3 buffers (fixed)Node: 3 * number_of_nodesCPU: 2.5 * number_of_cpus Here are some examples of what this means for different Node/CPU counts: Configuration None Node CPU 2 CPUs, 1 Node 3 buffers 3 buffers 5 buffers 6 CPUs, 2 Node 3 buffers 6 buffers 15 buffers 40 CPUs, 5 Nodes 3 buffers 15 buffers 100 buffers   Aside: Buffer size on multi-processor computersAs the number of Nodes or CPUs increases, the size of the event buffer gets smaller because the total memory is sub-divided into more pieces. The defaults will hold up to this for a while since each buffer set is holding events only from the Node or CPU that it is associated with, but at some point the buffers will get too small and you’ll either see events being dropped or you’ll get an error when you create your session because you’re below the minimum buffer size. Increase the MAX_MEMORY setting to an appropriate number for the configuration. The most likely reason to start partitioning is going to be related to performance. If you notice that running an event session is impacting the performance of your server beyond a reasonably expected level [Yes, there is a reasonably expected level of work required to collect events.] then partitioning might be an answer. Before you partition you might want to check a few other things: Is your event retention set to NO_EVENT_LOSS and causing blocking? (I told you not to do this.) Consider changing your event loss mode or increasing memory. Are you over collecting and causing more work than necessary? Consider adding predicates to events or removing unnecessary events and actions from your session. Are you writing the file target to the same slow disk that you use for TempDB and your other high activity databases? <kidding> <not really> It’s always worth considering the end to end picture – if you’re writing events to a file you can be impacted by I/O, network; all the usual stuff. Assuming you’ve ruled out the obvious (and not so obvious) issues, there are performance conditions that will be addressed by partitioning. For example, it’s possible to have a successful event session (eg. no dropped events) but still see a performance impact because you have many CPUs all attempting to write to the same free buffer and having to wait in line to finish their work. This is a case where partitioning would relieve the contention between the different CPUs and likely reduce the performance impact cause by the event session. There is no DMV you can check to find these conditions – sorry – that’s where the art comes in. This is  largely a matter of experimentation. On the bright side you probably won’t need to to worry about this level of detail all that often. The performance impact of Extended Events is significantly lower than what you may be used to with SQL Trace. You will likely only care about the impact if you are trying to set up a long running event session that will be part of your everyday workload – sessions used for short term troubleshooting will likely fall into the “reasonably expected impact” category. Hey buddy – I think you forgot something OK, there are two options I didn’t cover: STARTUP_STATE & TRACK_CAUSALITY. If you want your event sessions to start automatically when the server starts, set the STARTUP_STATE option to ON. (Now there is only one option I didn’t cover.) I’m going to leave causality for another post since it’s not really related to session behavior, it’s more about event analysis. - Mike Share this post: email it! | bookmark it! | digg it! | reddit! | kick it! | live it!

    Read the article

  • Oracle anuncia resultados de Q3 FY10

    - by Paulo Folgado
    Oracle Reports GAAP EPS of $0.23, Non-GAAP EPS of $0.38New Software Licenses Up 13%, Applications New Licenses Up 21%Oracle Corporation today announced fiscal 2010 Q3 GAAP total revenues were up 17% to $6.4 billion, while non-GAAP total revenues were up 18% to $6.5 billion. Excluding the impact of Sun Microsystems, Inc., which Oracle acquired on January 26, 2010, GAAP total revenue grew 7%. GAAP new software license revenues were up 13% to $1.7 billion, and up 10% to $1.7 billion excluding Sun. GAAP software license updates and product support revenues were up 13% to $3.3 billion, while non-GAAP software license updates and product support revenues were up 12% to $3.3 billion. GAAP operating income was down 5% to $1.8 billion, and GAAP operating margin was 29%. Non-GAAP operating income was up 13% to $2.9 billion, and non-GAAP operating margin was 45%. GAAP net income was down 10% to $1.2 billion, while non-GAAP net income was up 9% to $1.9 billion. GAAP earnings per share were $0.23, down 11% compared to last year while non-GAAP earnings per share were up 9% to $0.38. GAAP operating cash flow on a trailing twelve-month basis was $8.2 billion. "Our solid top line growth, coupled with disciplined expense management, was key in generating $8.0 billion of free cash flow over the last twelve months," said Oracle CFO Jeff Epstein."The Sun integration is going even better than we expected," said Oracle President, Safra Catz. "We believe that Sun will make a significant contribution to our fourth quarter earnings per share as well as meet the profitability goals we set for next year.""Exadata is the fastest growing product in Oracle's history," said Oracle President, Charles Phillips. "Introduced a little over a year ago, the Exadata pipeline is now approaching $400 million with Q4 bookings forecast at nearly $100 million. This strengthens both sales growth and profitability in our Sun server and storage businesses.""Every quarter we grab huge chunks of market share from SAP," said Oracle CEO, Larry Ellison. "SAP's most recent quarter was the best quarter of their year, only down 15%, while Oracle's application sales were up 21%. But SAP is well ahead of us in the number of CEOs for this year, announcing their third and fourth, while we only had one."In addition, Oracle's Board of Directors declared a cash dividend of $0.05 per share of outstanding common stock to be paid to stockholders of record as of the close of business on April 14, 2010, with a payment date of May 5, 2010. Future declarations of quarterly dividends and the establishment of future record and payment dates are subject to the final determination of Oracle's Board of Directors.Q3 Earnings Conference Call and WebcastOracle will hold a conference call and web broadcast today to discuss these results at 2:00 p.m. Pacific. You may listen to the call by dialing (800) 214-0694 or (719) 955-1425, Passcode: 567035. To access the live Web broadcast of this event, please visit the Oracle Investor Relations Web site at http://www.oracle.com/investor.

    Read the article

  • SQL – Building a High Traffic, Profitable Blog – A Unique Gift on Author’s Birthday

    - by Pinal Dave
    Every July 30th, I like to do something new. It is my birthday and I like to give gifts to everyone this day. Last year, at this time I had written an article A Year Older and 3 SQL Server Books and 3 Video Courses – 33. I had written a total of 3 books by that time and had published total of  3 Pluralsight courses. When I look back the year, I feel that I gave my best to last year. Sine Last July 30th, I have written 6 more books and 5 more video courses. The total is now 9 books and 8 video courses. It seems that I have been producing one new book or course every month since last July. Building a High Traffic, Profitable Blog Out of my 8 courses my favorite course is my latest course at Pluralsight. This course is about how to build a high traffic blog and monetize it. I have been blogging for over 7 years and there have been many hurdles and roadblocks but I have never stopped blogging any single day. There have been many instances when I felt I should just hit delete and remove my entire blog from the web but fortunately I had courage to stand by on my decisions. Well at the end, I kept on fighting through the difficult time and kept on blogging. Every day there was a lesson to learn and every day there was an issue to resolve. I never gave up and kept on building new content. Today after 7 years, when I look back there are many stories to tell. It was impossible to write down the stories so I decided to build a course based on my experience. In this course, I share all the best tricks to build a high traffic, profitable blog. When we talk about profit, people often talk about money but the reality is that profit is much bigger word than money. There are many different ways one can profit from their own blog. In this course, I discuss about all different ways about how you can be profitable by building a high traffic blog. I believe this course is for everybody who aspire to build a website or blog which gives them a profit.  Here are the major topics based out of this course. Introduction Techniques to Engage Blog Readers Social Media – Social Sharing and Social Networking Search Engine Optimization (SEO) Monetizing a Blog Frequently Asked Questions Checklists Personally I believe this is the best gift I can give to all of you my friends. Build a successful high traffic blog and monetize it. Here is the list of the all of my video courses and here is the list of all of the books. Reference: Pinal Dave (http://blog.sqlauthority.com) Filed under: About Me, PostADay, SQL, SQL Authority, SQL Query, SQL Server, SQL Tips and Tricks, T SQL Tagged: Blogging

    Read the article

  • curlftpfs mount disagrees with the fstab

    - by KayakJim
    I am working with curlftpfs to mount a remote FTP directory locally in Kubuntu 12.04 64-bit. I have the following entry in my /etc/fstab: curlftpfs#ftp_user:ftp_password@ftp_server /mnt/nimh fuse ro,noexec,nosuid,nodev,noauto,user,allow_other,uid=1000,gid=1000 0 0 I have created the directory in /mnt with the following: |-> ll /mnt total 4.0K drwxrwxr-x 2 jim fuse 4.0K Jan 6 09:56 nimh/ My user does belong to the fuse group as well: uid=1000(jim) gid=1000(jim) groups=1000(jim),27(sudo),105(fuse) I am able to mount manually without issue but then the /mnt changes to: |-> mount /mnt/nimh |-> ll /mnt total 0 drwxr-xr-x 1 jim jim 1.0K Dec 31 1969 nimh/ However when I attempt to umount /mnt/nimh I receive: umount: /mnt/nimh mount disagrees with the fstab My /etc/mtab looks like: curlftpfs#ftp://ftp_user:ftp_password@ftp_server/ /mnt/nimh fuse ro,noexec,nosuid,nodev,allow_other,user=jim 0 0 I am able to umount the filesystem without issue if I sudo. Any idea what I'm missing in order to be able to unmount without having to use sudo?

    Read the article

  • ASP.NET MVC 2 Model Binding for a Collection

    - by nmarun
    Yes, my yet another post on Model Binding (previous one is here), but this one uses features presented in MVC 2. How I got to writing this blog? Well, I’m on a project where we’re doing some MVC things for a shopping cart. Let me show you what I was working with. Below are my model classes: 1: public class Product 2: { 3: public int Id { get; set; } 4: public string Name { get; set; } 5: public int Quantity { get; set; } 6: public decimal UnitPrice { get; set; } 7: } 8:   9: public class Totals 10: { 11: public decimal SubTotal { get; set; } 12: public decimal Tax { get; set; } 13: public decimal Total { get; set; } 14: } 15:   16: public class Basket 17: { 18: public List<Product> Products { get; set; } 19: public Totals Totals { get; set;} 20: } The view looks as below:  1: <h2>Shopping Cart</h2> 2:   3: <% using(Html.BeginForm()) { %> 4: 5: <h3>Products</h3> 6: <% for (int i = 0; i < Model.Products.Count; i++) 7: { %> 8: <div style="width: 100px;float:left;">Id</div> 9: <div style="width: 100px;float:left;"> 10: <%= Html.TextBox("ID", Model.Products[i].Id) %> 11: </div> 12: <div style="clear:both;"></div> 13: <div style="width: 100px;float:left;">Name</div> 14: <div style="width: 100px;float:left;"> 15: <%= Html.TextBox("Name", Model.Products[i].Name) %> 16: </div> 17: <div style="clear:both;"></div> 18: <div style="width: 100px;float:left;">Quantity</div> 19: <div style="width: 100px;float:left;"> 20: <%= Html.TextBox("Quantity", Model.Products[i].Quantity)%> 21: </div> 22: <div style="clear:both;"></div> 23: <div style="width: 100px;float:left;">Unit Price</div> 24: <div style="width: 100px;float:left;"> 25: <%= Html.TextBox("UnitPrice", Model.Products[i].UnitPrice)%> 26: </div> 27: <div style="clear:both;"><hr /></div> 28: <% } %> 29: 30: <h3>Totals</h3> 31: <div style="width: 100px;float:left;">Sub Total</div> 32: <div style="width: 100px;float:left;"> 33: <%= Html.TextBox("SubTotal", Model.Totals.SubTotal)%> 34: </div> 35: <div style="clear:both;"></div> 36: <div style="width: 100px;float:left;">Tax</div> 37: <div style="width: 100px;float:left;"> 38: <%= Html.TextBox("Tax", Model.Totals.Tax)%> 39: </div> 40: <div style="clear:both;"></div> 41: <div style="width: 100px;float:left;">Total</div> 42: <div style="width: 100px;float:left;"> 43: <%= Html.TextBox("Total", Model.Totals.Total)%> 44: </div> 45: <div style="clear:both;"></div> 46: <p /> 47: <input type="submit" name="Submit" value="Submit" /> 48: <% } %> .csharpcode, .csharpcode pre { font-size: small; color: black; font-family: consolas, "Courier New", courier, monospace; background-color: #ffffff; /*white-space: pre;*/ } .csharpcode pre { margin: 0em; } .csharpcode .rem { color: #008000; } .csharpcode .kwrd { color: #0000ff; } .csharpcode .str { color: #006080; } .csharpcode .op { color: #0000c0; } .csharpcode .preproc { color: #cc6633; } .csharpcode .asp { background-color: #ffff00; } .csharpcode .html { color: #800000; } .csharpcode .attr { color: #ff0000; } .csharpcode .alt { background-color: #f4f4f4; width: 100%; margin: 0em; } .csharpcode .lnum { color: #606060; } .csharpcode, .csharpcode pre { font-size: small; color: black; font-family: consolas, "Courier New", courier, monospace; background-color: #ffffff; /*white-space: pre;*/ } .csharpcode pre { margin: 0em; } .csharpcode .rem { color: #008000; } .csharpcode .kwrd { color: #0000ff; } .csharpcode .str { color: #006080; } .csharpcode .op { color: #0000c0; } .csharpcode .preproc { color: #cc6633; } .csharpcode .asp { background-color: #ffff00; } .csharpcode .html { color: #800000; } .csharpcode .attr { color: #ff0000; } .csharpcode .alt { background-color: #f4f4f4; width: 100%; margin: 0em; } .csharpcode .lnum { color: #606060; } Nothing fancy, just a bunch of div’s containing textboxes and a submit button. Just make note that the textboxes have the same name as the property they are going to display. Yea, yea, I know. I’m displaying unit price as a textbox instead of a label, but that’s beside the point (and trust me, this will not be how it’ll look on the production site!!). The way my controller works is that initially two dummy products are added to the basked object and the Totals are calculated based on what products were added in what quantities and their respective unit price. So when the page loads in edit mode, where the user can change the quantity and hit the submit button. In the ‘post’ version of the action method, the Totals get recalculated and the new total will be displayed on the screen. Here’s the code: 1: public ActionResult Index() 2: { 3: Product product1 = new Product 4: { 5: Id = 1, 6: Name = "Product 1", 7: Quantity = 2, 8: UnitPrice = 200m 9: }; 10:   11: Product product2 = new Product 12: { 13: Id = 2, 14: Name = "Product 2", 15: Quantity = 1, 16: UnitPrice = 150m 17: }; 18:   19: List<Product> products = new List<Product> { product1, product2 }; 20:   21: Basket basket = new Basket 22: { 23: Products = products, 24: Totals = ComputeTotals(products) 25: }; 26: return View(basket); 27: } 28:   29: [HttpPost] 30: public ActionResult Index(Basket basket) 31: { 32: basket.Totals = ComputeTotals(basket.Products); 33: return View(basket); 34: } .csharpcode, .csharpcode pre { font-size: small; color: black; font-family: consolas, "Courier New", courier, monospace; background-color: #ffffff; /*white-space: pre;*/ } .csharpcode pre { margin: 0em; } .csharpcode .rem { color: #008000; } .csharpcode .kwrd { color: #0000ff; } .csharpcode .str { color: #006080; } .csharpcode .op { color: #0000c0; } .csharpcode .preproc { color: #cc6633; } .csharpcode .asp { background-color: #ffff00; } .csharpcode .html { color: #800000; } .csharpcode .attr { color: #ff0000; } .csharpcode .alt { background-color: #f4f4f4; width: 100%; margin: 0em; } .csharpcode .lnum { color: #606060; } That’s that. Now I run the app, I see two products with the totals section below them. I look at the view source and I see that the input controls have the right ID, the right name and the right value as well. 1: <input id="ID" name="ID" type="text" value="1" /> 2: <input id="Name" name="Name" type="text" value="Product 1" /> 3: ... 4: <input id="ID" name="ID" type="text" value="2" /> 5: <input id="Name" name="Name" type="text" value="Product 2" /> .csharpcode, .csharpcode pre { font-size: small; color: black; font-family: consolas, "Courier New", courier, monospace; background-color: #ffffff; /*white-space: pre;*/ } .csharpcode pre { margin: 0em; } .csharpcode .rem { color: #008000; } .csharpcode .kwrd { color: #0000ff; } .csharpcode .str { color: #006080; } .csharpcode .op { color: #0000c0; } .csharpcode .preproc { color: #cc6633; } .csharpcode .asp { background-color: #ffff00; } .csharpcode .html { color: #800000; } .csharpcode .attr { color: #ff0000; } .csharpcode .alt { background-color: #f4f4f4; width: 100%; margin: 0em; } .csharpcode .lnum { color: #606060; } So just as a regular user would do, I change the quantity value of one of the products and hit the submit button. The ‘post’ version of the Index method gets called and I had put a break-point on line 32 in the above snippet. When I hovered my mouse on the ‘basked’ object, happily assuming that the object would be all bound and ready for use, I was surprised to see both basket.Products and basket.Totals were null. Huh? A little research and I found out that the reason the DefaultModelBinder could not do its job is because of a naming mismatch on the input controls. What I mean is that when you have to bind to a custom .net type, you need more than just the property name. You need to pass a qualified name to the name property of the input control. I modified my view and the emitted code looked as below: 1: <input id="Product_Name" name="Product.Name" type="text" value="Product 1" /> 2: ... 3: <input id="Product_Name" name="Product.Name" type="text" value="Product 2" /> 4: ... 5: <input id="Totals_SubTotal" name="Totals.SubTotal" type="text" value="550" /> .csharpcode, .csharpcode pre { font-size: small; color: black; font-family: consolas, "Courier New", courier, monospace; background-color: #ffffff; /*white-space: pre;*/ } .csharpcode pre { margin: 0em; } .csharpcode .rem { color: #008000; } .csharpcode .kwrd { color: #0000ff; } .csharpcode .str { color: #006080; } .csharpcode .op { color: #0000c0; } .csharpcode .preproc { color: #cc6633; } .csharpcode .asp { background-color: #ffff00; } .csharpcode .html { color: #800000; } .csharpcode .attr { color: #ff0000; } .csharpcode .alt { background-color: #f4f4f4; width: 100%; margin: 0em; } .csharpcode .lnum { color: #606060; } Now, I update the quantity and hit the submit button and I see that the Totals object is populated, but the Products list is still null. Once again I went: ‘Hmm.. time for more research’. I found out that the way to do this is to provide the name as: 1: <%= Html.TextBox(string.Format("Products[{0}].ID", i), Model.Products[i].Id) %> 2: <!-- this will be rendered as --> 3: <input id="Products_0__ID" name="Products[0].ID" type="text" value="1" /> .csharpcode, .csharpcode pre { font-size: small; color: black; font-family: consolas, "Courier New", courier, monospace; background-color: #ffffff; /*white-space: pre;*/ } .csharpcode pre { margin: 0em; } .csharpcode .rem { color: #008000; } .csharpcode .kwrd { color: #0000ff; } .csharpcode .str { color: #006080; } .csharpcode .op { color: #0000c0; } .csharpcode .preproc { color: #cc6633; } .csharpcode .asp { background-color: #ffff00; } .csharpcode .html { color: #800000; } .csharpcode .attr { color: #ff0000; } .csharpcode .alt { background-color: #f4f4f4; width: 100%; margin: 0em; } .csharpcode .lnum { color: #606060; } It was only now that I was able to see both the products and the totals being properly bound in the ‘post’ action method. Somehow, I feel this is kinda ‘clunky’ way of doing things. Seems like people at MS felt in a similar way and offered us a much cleaner way to solve this issue. The simple solution is that instead of using a Textbox, we can either use a TextboxFor or an EditorFor helper method. This one directly spits out the name of the input property as ‘Products[0].ID and so on. Cool right? I totally fell for this and changed my UI to contain EditorFor helper method. At this point, I ran the application, changed the quantity field and pressed the submit button. Of course my basket object parameter in my action method was correctly bound after these changes. I let the app complete the rest of the lines in the action method. When the page finally rendered, I did see that the quantity was changed to what I entered before the post. But, wait a minute, the totals section did not reflect the changes and showed the old values. My status: COMPLETELY PUZZLED! Just to recap, this is what my ‘post’ Index method looked like: 1: [HttpPost] 2: public ActionResult Index(Basket basket) 3: { 4: basket.Totals = ComputeTotals(basket.Products); 5: return View(basket); 6: } .csharpcode, .csharpcode pre { font-size: small; color: black; font-family: consolas, "Courier New", courier, monospace; background-color: #ffffff; /*white-space: pre;*/ } .csharpcode pre { margin: 0em; } .csharpcode .rem { color: #008000; } .csharpcode .kwrd { color: #0000ff; } .csharpcode .str { color: #006080; } .csharpcode .op { color: #0000c0; } .csharpcode .preproc { color: #cc6633; } .csharpcode .asp { background-color: #ffff00; } .csharpcode .html { color: #800000; } .csharpcode .attr { color: #ff0000; } .csharpcode .alt { background-color: #f4f4f4; width: 100%; margin: 0em; } .csharpcode .lnum { color: #606060; } A careful debug confirmed that the basked.Products[0].Quantity showed the updated value and the ComputeTotals() method also returns the correct totals. But still when I passed this basket object, it ended up showing the old totals values only. I began playing a bit with the code and my first guess was that the input controls got their values from the ModelState object. For those who don’t know, the ModelState is a temporary storage area that ASP.NET MVC uses to retain incoming attempted values plus binding and validation errors. Also, the fact that input controls populate the values using data taken from: Previously attempted values recorded in the ModelState["name"].Value.AttemptedValue Explicitly provided value (<%= Html.TextBox("name", "Some value") %>) ViewData, by calling ViewData.Eval("name") FYI: ViewData dictionary takes precedence over ViewData's Model properties – read more here. These two indicators led to my guess. It took me quite some time, but finally I hit this post where Brad brilliantly explains why this is the preferred behavior. My guess was right and I, accordingly modified my code to reflect the following way: 1: [HttpPost] 2: public ActionResult Index(Basket basket) 3: { 4: // read the following posts to see why the ModelState 5: // needs to be cleared before passing it the view 6: // http://forums.asp.net/t/1535846.aspx 7: // http://forums.asp.net/p/1527149/3687407.aspx 8: if (ModelState.IsValid) 9: { 10: ModelState.Clear(); 11: } 12:   13: basket.Totals = ComputeTotals(basket.Products); 14: return View(basket); 15: } .csharpcode, .csharpcode pre { font-size: small; color: black; font-family: consolas, "Courier New", courier, monospace; background-color: #ffffff; /*white-space: pre;*/ } .csharpcode pre { margin: 0em; } .csharpcode .rem { color: #008000; } .csharpcode .kwrd { color: #0000ff; } .csharpcode .str { color: #006080; } .csharpcode .op { color: #0000c0; } .csharpcode .preproc { color: #cc6633; } .csharpcode .asp { background-color: #ffff00; } .csharpcode .html { color: #800000; } .csharpcode .attr { color: #ff0000; } .csharpcode .alt { background-color: #f4f4f4; width: 100%; margin: 0em; } .csharpcode .lnum { color: #606060; } What this does is that in the case where your ModelState IS valid, it clears the dictionary. This enables the values to be read from the model directly and not from the ModelState. So the verdict is this: If you need to pass other parameters (like html attributes and the like) to your input control, use 1: <%= Html.TextBox(string.Format("Products[{0}].ID", i), Model.Products[i].Id) %> .csharpcode, .csharpcode pre { font-size: small; color: black; font-family: consolas, "Courier New", courier, monospace; background-color: #ffffff; /*white-space: pre;*/ } .csharpcode pre { margin: 0em; } .csharpcode .rem { color: #008000; } .csharpcode .kwrd { color: #0000ff; } .csharpcode .str { color: #006080; } .csharpcode .op { color: #0000c0; } .csharpcode .preproc { color: #cc6633; } .csharpcode .asp { background-color: #ffff00; } .csharpcode .html { color: #800000; } .csharpcode .attr { color: #ff0000; } .csharpcode .alt { background-color: #f4f4f4; width: 100%; margin: 0em; } .csharpcode .lnum { color: #606060; } Since, in EditorFor, there is no direct and simple way of passing this information to the input control. If you don’t have to pass any such ‘extra’ piece of information to the control, then go the EditorFor way. The code used in the post can be found here.

    Read the article

  • SQL SERVER – SSMS: Top Object and Batch Execution Statistics Reports

    - by Pinal Dave
    The month of June till mid of July has been the fever of sports. First, it was Wimbledon Tennis and then the Soccer fever was all over. There is a huge number of fan followers and it is great to see the level at which people sometimes worship these sports. Being an Indian, I cannot forget to mention the India tour of England later part of July. Following these sports and as the events unfold to the finals, there are a number of ways the statisticians can slice and dice the numbers. Cue from soccer I can surely say there is a team performance against another team and then there is individual member fairs against a particular opponent. Such statistics give us a fair idea to how a team in the past or in the recent past has fared against each other, head-to-head stats during World cup and during other neutral venue games. All these statistics are just pointers. In reality, they don’t reflect the calibre of the current team because the individuals who performed in each of these games are totally different (Typical example being the Brazil Vs Germany semi-final match in FIFA 2014). So at times these numbers are misleading. It is worth investigating and get the next level information. Similar to these statistics, SQL Server Management studio is also equipped with a number of reports like a) Object Execution Statistics report and b) Batch Execution Statistics reports. As discussed in the example, the team scorecard is like the Batch Execution statistics and individual stats is like Object Level statistics. The analogy can be taken only this far, trust me there is no correlation between SQL Server functioning and playing sports – It is like I think about diet all the time except while I am eating. Performance – Batch Execution Statistics Let us view the first report which can be invoked from Server Node -> Reports -> Standard Reports -> Performance – Batch Execution Statistics. Most of the values that are displayed in this report come from the DMVs sys.dm_exec_query_stats and sys.dm_exec_sql_text(sql_handle). This report contains 3 distinctive sections as outline below.   Section 1: This is a graphical bar graph representation of Average CPU Time, Average Logical reads and Average Logical Writes for individual batches. The Batch numbers are indicative and the details of individual batch is available in section 3 (detailed below). Section 2: This represents a Pie chart of all the batches by Total CPU Time (%) and Total Logical IO (%) by batches. This graphical representation tells us which batch consumed the highest CPU and IO since the server started, provided plan is available in the cache. Section 3: This is the section where we can find the SQL statements associated with each of the batch Numbers. This also gives us the details of Average CPU / Average Logical Reads and Average Logical Writes in the system for the given batch with object details. Expanding the rows, I will also get the # Executions and # Plans Generated for each of the queries. Performance – Object Execution Statistics The second report worth a look is Object Execution statistics. This is a similar report as the previous but turned on its head by SQL Server Objects. The report has 3 areas to look as above. Section 1 gives the Average CPU, Average IO bar charts for specific objects. The section 2 is a graphical representation of Total CPU by objects and Total Logical IO by objects. The final section details the various objects in detail with the Avg. CPU, IO and other details which are self-explanatory. At a high-level both the reports are based on queries on two DMVs (sys.dm_exec_query_stats and sys.dm_exec_sql_text) and it builds values based on calculations using columns in them: SELECT * FROM    sys.dm_exec_query_stats s1 CROSS APPLY sys.dm_exec_sql_text(sql_handle) AS s2 WHERE   s2.objectid IS NOT NULL AND DB_NAME(s2.dbid) IS NOT NULL ORDER BY  s1.sql_handle; This is one of the simplest form of reports and in future blogs we will look at more complex reports. I truly hope that these reports can give DBAs and developers a hint about what is the possible performance tuning area. As a closing point I must emphasize that all above reports pick up data from the plan cache. If a particular query has consumed a lot of resources earlier, but plan is not available in the cache, none of the above reports would show that bad query. Reference: Pinal Dave (http://blog.sqlauthority.com)Filed under: SQL, SQL Authority, SQL Query, SQL Server, SQL Server Management Studio, SQL Tips and Tricks, T SQL Tagged: SQL Reports

    Read the article

  • Mysql performance problem & Failed DIMM

    - by murdoch
    Hi I have a dedicated mysql database server which has been having some performance problems recently, under normal load the server will be running fine, then suddenly out of the blue the performance will fall off a cliff. The server isn't using the swap file and there is 12GB of RAM in the server, more than enough for its needs. After contacting my hosting comapnies support they have discovered that there is a failed 2GB DIMM in the server and have scheduled to replace it tomorow morning. My question is could a failed DIMM result in the performance problems I am seeing or is this just coincidence? My worry is that they will replace the ram tomorrow but the problems will persist and I will still be lost of explanations so I am just trying to think ahead. The reason I ask is that there is plenty of RAM in the server, more than required and simply missing 2GB should be a problem, so if this failed DIMM is causing these performance problems then the OS must be trying to access the failed DIMM and slowing down as a result. Does that sound like a credible explanation? This is what DELLs omreport program says about the RAM, notice one dimm is "Critical" Memory Information Health : Critical Memory Operating Mode Fail Over State : Inactive Memory Operating Mode Configuration : Optimizer Attributes of Memory Array(s) Attributes : Location Memory Array 1 : System Board or Motherboard Attributes : Use Memory Array 1 : System Memory Attributes : Installed Capacity Memory Array 1 : 12288 MB Attributes : Maximum Capacity Memory Array 1 : 196608 MB Attributes : Slots Available Memory Array 1 : 18 Attributes : Slots Used Memory Array 1 : 6 Attributes : ECC Type Memory Array 1 : Multibit ECC Total of Memory Array(s) Attributes : Total Installed Capacity Value : 12288 MB Attributes : Total Installed Capacity Available to the OS Value : 12004 MB Attributes : Total Maximum Capacity Value : 196608 MB Details of Memory Array 1 Index : 0 Status : Ok Connector Name : DIMM_A1 Type : DDR3-Registered Size : 2048 MB Index : 1 Status : Ok Connector Name : DIMM_A2 Type : DDR3-Registered Size : 2048 MB Index : 2 Status : Ok Connector Name : DIMM_A3 Type : DDR3-Registered Size : 2048 MB Index : 3 Status : Critical Connector Name : DIMM_B1 Type : DDR3-Registered Size : 2048 MB Index : 4 Status : Ok Connector Name : DIMM_B2 Type : DDR3-Registered Size : 2048 MB Index : 5 Status : Ok Connector Name : DIMM_B3 Type : DDR3-Registered Size : 2048 MB the command free -m shows this, the server seems to be using more than 10GB of ram which would suggest it is trying to use the DIMM total used free shared buffers cached Mem: 12004 10766 1238 0 384 4809 -/+ buffers/cache: 5572 6432 Swap: 2047 0 2047 iostat output while problem is occuring avg-cpu: %user %nice %system %iowait %steal %idle 52.82 0.00 11.01 0.00 0.00 36.17 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 47.00 0.00 576.00 0 576 sda1 0.00 0.00 0.00 0 0 sda2 1.00 0.00 32.00 0 32 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 46.00 0.00 544.00 0 544 avg-cpu: %user %nice %system %iowait %steal %idle 53.12 0.00 7.81 0.00 0.00 39.06 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 49.00 0.00 592.00 0 592 sda1 0.00 0.00 0.00 0 0 sda2 0.00 0.00 0.00 0 0 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 49.00 0.00 592.00 0 592 avg-cpu: %user %nice %system %iowait %steal %idle 56.09 0.00 7.43 0.37 0.00 36.10 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 232.00 0.00 64520.00 0 64520 sda1 0.00 0.00 0.00 0 0 sda2 159.00 0.00 63728.00 0 63728 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 73.00 0.00 792.00 0 792 avg-cpu: %user %nice %system %iowait %steal %idle 52.18 0.00 9.24 0.06 0.00 38.51 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 49.00 0.00 600.00 0 600 sda1 0.00 0.00 0.00 0 0 sda2 0.00 0.00 0.00 0 0 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 49.00 0.00 600.00 0 600 avg-cpu: %user %nice %system %iowait %steal %idle 54.82 0.00 8.64 0.00 0.00 36.55 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 100.00 0.00 2168.00 0 2168 sda1 0.00 0.00 0.00 0 0 sda2 0.00 0.00 0.00 0 0 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 100.00 0.00 2168.00 0 2168 avg-cpu: %user %nice %system %iowait %steal %idle 54.78 0.00 6.75 0.00 0.00 38.48 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 84.00 0.00 896.00 0 896 sda1 0.00 0.00 0.00 0 0 sda2 0.00 0.00 0.00 0 0 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 84.00 0.00 896.00 0 896 avg-cpu: %user %nice %system %iowait %steal %idle 54.34 0.00 7.31 0.00 0.00 38.35 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 81.00 0.00 840.00 0 840 sda1 0.00 0.00 0.00 0 0 sda2 0.00 0.00 0.00 0 0 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 81.00 0.00 840.00 0 840 avg-cpu: %user %nice %system %iowait %steal %idle 55.18 0.00 5.81 0.44 0.00 38.58 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 317.00 0.00 105632.00 0 105632 sda1 0.00 0.00 0.00 0 0 sda2 224.00 0.00 104672.00 0 104672 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 93.00 0.00 960.00 0 960 avg-cpu: %user %nice %system %iowait %steal %idle 55.38 0.00 7.63 0.00 0.00 36.98 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 74.00 0.00 800.00 0 800 sda1 0.00 0.00 0.00 0 0 sda2 0.00 0.00 0.00 0 0 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 74.00 0.00 800.00 0 800 avg-cpu: %user %nice %system %iowait %steal %idle 56.43 0.00 7.80 0.00 0.00 35.77 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 72.00 0.00 784.00 0 784 sda1 0.00 0.00 0.00 0 0 sda2 0.00 0.00 0.00 0 0 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 72.00 0.00 784.00 0 784 avg-cpu: %user %nice %system %iowait %steal %idle 54.87 0.00 6.49 0.00 0.00 38.64 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 80.20 0.00 855.45 0 864 sda1 0.00 0.00 0.00 0 0 sda2 0.00 0.00 0.00 0 0 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 80.20 0.00 855.45 0 864 avg-cpu: %user %nice %system %iowait %steal %idle 57.22 0.00 5.69 0.00 0.00 37.09 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 33.00 0.00 432.00 0 432 sda1 0.00 0.00 0.00 0 0 sda2 0.00 0.00 0.00 0 0 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 33.00 0.00 432.00 0 432 avg-cpu: %user %nice %system %iowait %steal %idle 56.03 0.00 7.93 0.00 0.00 36.04 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 41.00 0.00 560.00 0 560 sda1 0.00 0.00 0.00 0 0 sda2 2.00 0.00 88.00 0 88 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 39.00 0.00 472.00 0 472 avg-cpu: %user %nice %system %iowait %steal %idle 55.78 0.00 5.13 0.00 0.00 39.09 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 29.00 0.00 392.00 0 392 sda1 0.00 0.00 0.00 0 0 sda2 0.00 0.00 0.00 0 0 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 29.00 0.00 392.00 0 392 avg-cpu: %user %nice %system %iowait %steal %idle 53.68 0.00 8.30 0.06 0.00 37.95 Device: tps Blk_read/s Blk_wrtn/s Blk_read Blk_wrtn sda 78.00 0.00 4280.00 0 4280 sda1 0.00 0.00 0.00 0 0 sda2 0.00 0.00 0.00 0 0 sda3 0.00 0.00 0.00 0 0 sda4 0.00 0.00 0.00 0 0 sda5 78.00 0.00 4280.00 0 4280

    Read the article

  • Ranking Part III

    - by PointsToShare
    © 2011 By: Dov Trietsch. All rights reserved   Ranking Part III In a previous blogs “Ranking an Introduction” and  “Ranking Part II” , you have already praised me in “Rank the Author” and learned how to create a new element on a page and how to place it where you need it. For this installment, I just added code to keep the number of votes (you vote by clicking one of the stars) and the total vote. Using these two, we can compute the average rating. It’s a small step, but its purpose is to show that we do not need a detailed history in order to compute the average. A running total is sufficient. Please note that once you close the game, you will lose your previous total. In real life, we persist the totals in the list itself. We also keep a list of actual votes, but its purpose is to prevent double votes. If a person has already voted, his user id is already on the list and our program will check for it and bar the person from voting again. This is coded in an event receiver, which is a SharePoint server piece of code. I will show you how to do this part in a subsequent blog. Again, go to the page and look at the code. The gist of it is here. avg, votes, and stars are global variables that I defined before. function sendRate(sel){//I hate long line so I created pieces of the message in their own vars            var s1 = "Your Rating Was: ";            var s2 = ".. ";            var s3 = "\nVotes = ";            var s4 = "\nTotal Stars = ";            var s5 = "\nAverage = ";            var s;            s = parseInt(sel.id.replace("_", '')); // Get the selected star number            votes = parseInt(votes) + 1;            stars = parseInt(stars) + s;            avg = parseFloat(stars) / parseFloat(votes);            alert(s1 + sel.id + s2 +sel.title + s3 + votes + s4 + stars + s5 + avg);} Click on the link to play and examine “Ranking with Stats” That’s all folks!

    Read the article

  • Ranking - an Introduction

    - by PointsToShare
    © 2011 By: Dov Trietsch. All rights reserved Ranking Ranking is quite common in the internet. Readers are asked to rank their latest reading by clicking on one of 5 (sometimes 10) stars. The number of stars is then converted to a number and the average number of stars as selected by all the readers is proudly (or shamefully) displayed for future readers. SharePoint 2007 lacked this feature altogether. SharePoint 2010 allows the users to rank items in a list or documents in a library (the two are actually the same because a library is actually a list). But in SP2010 the computation of the average is done later on a timer rather than on-the-spot as it should be. I suspect that the reason for this shortcoming is that they did not involve a mathematician! Let me explain. Ranking is kept in a related list. When a user rates a document the rank-list is added an item with the item id, the user name, and his number of stars. The fact that a user already ranked an item prevents him from ranking it again. This prevents the creator of the item from asking his mother to rank it a 5 and do it 753 times, thus stacking the ballot. Some systems will allow a user to change his rating and this will be done by updating the rank-list item. Now, when the timer kicks off, the list is spanned and for each item the rank-list items containing this id are summed up and divided by the number of votes thus yielding the new average. This is obviously very time consuming and very server intensive. In the 18th century an early actuary named James Dodson used what the great Augustus De Morgan (of De Morgan’s law) later named Commutation tables. The labor involved in computing a life insurance premium was staggering and also very error prone. Clerks with pencil and paper would multiply and add mountains of numbers to do the task. The more steps the greater the probability of error and the more expensive the process. Commutation tables created a “summary” of many steps and reduced the work 100 fold. So had Microsoft taken a lesson in the history of computation, they would have developed a much faster way for rating that may be done in real-time and is also 100 times faster and less CPU intensive. How do we do this? We use a form of commutation. We always keep the number of votes and the total of stars. One simple division gives us the average. So we write an event receiver. When a vote is added, we just add the stars to the total-stars and 1 to the number of votes. We then recomputed the average. When a vote is updated, we reduce the total by the old vote, increase it by the new vote and leave the number of votes the same. Again we do the division to get the new average. When a vote is deleted (highly unlikely and maybe even prohibited), we reduce the total by that vote and reduce the number of votes by 1… Gone are the days of spanning lists, counting items, and tallying votes and we have no need for a timer process to run it all. This is the first of a few treatises on ranking. Even though I discussed the math and the history thereof, in here I am only going to solve the presentation issue. I wanted to create the CSS and Jscript needed to display the stars, create the various effects like hovering and clicking (onmouseover, onmouseout, onclick, etc.) and I wanted to create a general solution with any number of stars. When I had it all done, I created the ranking game so that I could test it. The game is interesting in and on itself, so here it is (or go to the games page and select “rank the stars”). BTW, when you play it, look at the source code and see how it was all done.  Next, how the 5 stars are displayed in the New and Update forms. When the whole set of articles will be done, you’ll be able to create the complete solution. That’s all folks!

    Read the article

  • Cannot add repository key

    - by William Anthony
    I just installed my new laptop with ubuntu 12.04 and when I'm trying to add key, there is a "network unreachable" error. william@ubuntu:~$ gpg --keyserver hkp://keys.gnupg.net --recv-keys 1C4CBDCDCD2EFD2A gpg: requesting key CD2EFD2A from hkp server keys.gnupg.net ?: keys.gnupg.net: Network is unreachable gpgkeys: HTTP fetch error 7: couldn't connect: Network is unreachable gpg: no valid OpenPGP data found. gpg: Total number processed: 0 I'm so sure the keyserver is not down, because I tried it again at my old laptop running ubuntu 11.04 william@william:~$ gpg --keyserver hkp://keys.gnupg.net --recv-keys 1C4CBDCDCD2EFD2A gpg: requesting key CD2EFD2A from hkp server keys.gnupg.net gpg: key CD2EFD2A: "Percona MySQL Development Team <[email protected]>" not changed gpg: Total number processed: 1 gpg: unchanged: 1 Is this a bug?

    Read the article

  • Experience formula with javascript

    - by StealingMana
    I'm having trouble working out a formula using this experience curve to get the total exp after each level. I bet its easy and im just over thinking it. maxlvl = 10; increment = 28; baseexp = 100; function calc(){ for (i = 0;i<(maxlvl*increment);i+=increment){ expperlvl = baseexp + i; document.writeln(expperlvl); } } I figured it out. maxlvl=6; base=200; increment=56; function total(){ totalxp= (base*(maxlvl-1))+(increment*(maxlvl-2)*(maxlvl-1)/2); document.write(totalxp); }

    Read the article

  • Data structure: sort and search effectively

    - by Jiten Shah
    I need to have a data structure with say 4 keys . I can sort on any of these keys. What data structure can I opt for? Sorting time should be very little. I thought of a tree, but it will be only help searching on one key. For other keys I'll have to remake the tree on that particular key and then find it. Is there any data structure that can take care of all 4 keys at the same time? these 4 fields are of total 12 bytes and total size for each record - 40 bytes.. have memory constraints too... operations are : insertion, deletion, sorting on different keys.

    Read the article

  • Server Memory with Magento

    - by Mohamed Elgharabawy
    I have a cloud server with the following specifications: 2vCPUs 4G RAM 160GB Disk Space Network 400Mb/s System Image: Ubuntu 12.04 LTS I am only running Magento CE 1.7.0.2 on this server. Nothing else. Usually, the server has a loading time of 4-5 seconds. Recently, this has dropped to over 30 seconds and sometimes the server just goes away and I get HTTP error reports to my email stating that HTTP requests took more than 20000ms. Running top command and sorting them returns the following: top - 15:29:07 up 3:40, 1 user, load average: 28.59, 25.95, 22.91 Tasks: 112 total, 30 running, 82 sleeping, 0 stopped, 0 zombie Cpu(s): 90.2%us, 9.3%sy, 0.0%ni, 0.0%id, 0.0%wa, 0.0%hi, 0.3%si, 0.2%st PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 31901 www-data 20 0 360m 71m 5840 R 7 1.8 1:39.51 apache2 32084 www-data 20 0 362m 72m 5548 R 7 1.8 1:31.56 apache2 32089 www-data 20 0 348m 59m 5660 R 7 1.5 1:41.74 apache2 32295 www-data 20 0 343m 54m 5532 R 7 1.4 2:00.78 apache2 32303 www-data 20 0 354m 65m 5260 R 7 1.6 1:38.76 apache2 32304 www-data 20 0 346m 56m 5544 R 7 1.4 1:41.26 apache2 32305 www-data 20 0 348m 59m 5640 R 7 1.5 1:50.11 apache2 32291 www-data 20 0 358m 69m 5256 R 6 1.7 1:44.26 apache2 32517 www-data 20 0 345m 56m 5532 R 6 1.4 1:45.56 apache2 30473 www-data 20 0 355m 66m 5680 R 6 1.7 2:00.05 apache2 32093 www-data 20 0 352m 63m 5848 R 6 1.6 1:53.23 apache2 32302 www-data 20 0 345m 56m 5512 R 6 1.4 1:55.87 apache2 32433 www-data 20 0 346m 57m 5500 S 6 1.4 1:31.58 apache2 32638 www-data 20 0 354m 65m 5508 R 6 1.6 1:36.59 apache2 32230 www-data 20 0 347m 57m 5524 R 6 1.4 1:33.96 apache2 32231 www-data 20 0 355m 66m 5512 R 6 1.7 1:37.47 apache2 32233 www-data 20 0 354m 64m 6032 R 6 1.6 1:59.74 apache2 32300 www-data 20 0 355m 66m 5672 R 6 1.7 1:43.76 apache2 32510 www-data 20 0 347m 58m 5512 R 6 1.5 1:42.54 apache2 32521 www-data 20 0 348m 59m 5508 R 6 1.5 1:47.99 apache2 32639 www-data 20 0 344m 55m 5512 R 6 1.4 1:34.25 apache2 32083 www-data 20 0 345m 56m 5696 R 5 1.4 1:59.42 apache2 32085 www-data 20 0 347m 58m 5692 R 5 1.5 1:42.29 apache2 32293 www-data 20 0 353m 64m 5676 R 5 1.6 1:52.73 apache2 32301 www-data 20 0 348m 59m 5564 R 5 1.5 1:49.63 apache2 32528 www-data 20 0 351m 62m 5520 R 5 1.6 1:36.11 apache2 31523 mysql 20 0 3460m 576m 8288 S 5 14.4 2:06.91 mysqld 32002 www-data 20 0 345m 55m 5512 R 5 1.4 2:01.88 apache2 32080 www-data 20 0 357m 68m 5512 S 5 1.7 1:31.30 apache2 32163 www-data 20 0 347m 58m 5512 S 5 1.5 1:58.68 apache2 32509 www-data 20 0 345m 56m 5504 R 5 1.4 1:49.54 apache2 32306 www-data 20 0 358m 68m 5504 S 4 1.7 1:53.29 apache2 32165 www-data 20 0 344m 55m 5524 S 4 1.4 1:40.71 apache2 32640 www-data 20 0 345m 56m 5528 R 4 1.4 1:36.49 apache2 31888 www-data 20 0 359m 70m 5664 R 4 1.8 1:57.07 apache2 32511 www-data 20 0 357m 67m 5512 S 3 1.7 1:47.00 apache2 32054 www-data 20 0 357m 68m 5660 S 2 1.7 1:53.10 apache2 1 root 20 0 24452 2276 1232 S 0 0.1 0:01.58 init Moreover, running free -m returns the following: total used free shared buffers cached Mem: 4003 3919 83 0 118 901 -/+ buffers/cache: 2899 1103 Swap: 0 0 0 To investigate this further, I have installed apache buddy, it recommeneded that I need to reduce the maxclient connections. Which I did. I also installed MysqlTuner and it suggests that I need to set my innodb_buffer_pool_size to = 3.0G. However, I cannot do that, since the whole memory is 4G. Here is the output from apache buddy: ### GENERAL REPORT ### Settings considered for this report: Your server's physical RAM: 4003MB Apache's MaxClients directive: 40 Apache MPM Model: prefork Largest Apache process (by memory): 73.77MB [ OK ] Your MaxClients setting is within an acceptable range. Max potential memory usage: 2950.8 MB Percentage of RAM allocated to Apache 73.72 % And this is the output of MySQLTuner: -------- Performance Metrics ------------------------------------------------- [--] Up for: 47m 22s (675K q [237.552 qps], 12K conn, TX: 1B, RX: 300M) [--] Reads / Writes: 45% / 55% [--] Total buffers: 2.1G global + 2.7M per thread (151 max threads) [OK] Maximum possible memory usage: 2.5G (64% of installed RAM) [OK] Slow queries: 0% (0/675K) [OK] Highest usage of available connections: 26% (40/151) [OK] Key buffer size / total MyISAM indexes: 36.0M/18.7M [OK] Key buffer hit rate: 100.0% (245K cached / 105 reads) [OK] Query cache efficiency: 92.5% (500K cached / 541K selects) [!!] Query cache prunes per day: 302886 [OK] Sorts requiring temporary tables: 0% (1 temp sorts / 15K sorts) [!!] Joins performed without indexes: 12135 [OK] Temporary tables created on disk: 25% (8K on disk / 32K total) [OK] Thread cache hit rate: 90% (1K created / 12K connections) [!!] Table cache hit rate: 17% (400 open / 2K opened) [OK] Open file limit used: 12% (123/1K) [OK] Table locks acquired immediately: 100% (196K immediate / 196K locks) [!!] InnoDB buffer pool / data size: 2.0G/3.5G [OK] InnoDB log waits: 0 -------- Recommendations ----------------------------------------------------- General recommendations: Run OPTIMIZE TABLE to defragment tables for better performance MySQL started within last 24 hours - recommendations may be inaccurate Enable the slow query log to troubleshoot bad queries Adjust your join queries to always utilize indexes Increase table_cache gradually to avoid file descriptor limits Read this before increasing table_cache over 64: http://bit.ly/1mi7c4C Variables to adjust: query_cache_size ( 64M) join_buffer_size ( 128.0K, or always use indexes with joins) table_cache ( 400) innodb_buffer_pool_size (= 3G) Last but not least, the server still has more than 60% of free disk space. Now, based on the above, I have few questions: Are these numbers normal? Do they make sense? Do I need to upgrade the server? If I don't need to upgrade and my configuration is not correct, how do I optimize it?

    Read the article

  • PTLQueue : a scalable bounded-capacity MPMC queue

    - by Dave
    Title: Fast concurrent MPMC queue -- I've used the following concurrent queue algorithm enough that it warrants a blog entry. I'll sketch out the design of a fast and scalable multiple-producer multiple-consumer (MPSC) concurrent queue called PTLQueue. The queue has bounded capacity and is implemented via a circular array. Bounded capacity can be a useful property if there's a mismatch between producer rates and consumer rates where an unbounded queue might otherwise result in excessive memory consumption by virtue of the container nodes that -- in some queue implementations -- are used to hold values. A bounded-capacity queue can provide flow control between components. Beware, however, that bounded collections can also result in resource deadlock if abused. The put() and take() operators are partial and wait for the collection to become non-full or non-empty, respectively. Put() and take() do not allocate memory, and are not vulnerable to the ABA pathologies. The PTLQueue algorithm can be implemented equally well in C/C++ and Java. Partial operators are often more convenient than total methods. In many use cases if the preconditions aren't met, there's nothing else useful the thread can do, so it may as well wait via a partial method. An exception is in the case of work-stealing queues where a thief might scan a set of queues from which it could potentially steal. Total methods return ASAP with a success-failure indication. (It's tempting to describe a queue or API as blocking or non-blocking instead of partial or total, but non-blocking is already an overloaded concurrency term. Perhaps waiting/non-waiting or patient/impatient might be better terms). It's also trivial to construct partial operators by busy-waiting via total operators, but such constructs may be less efficient than an operator explicitly and intentionally designed to wait. A PTLQueue instance contains an array of slots, where each slot has volatile Turn and MailBox fields. The array has power-of-two length allowing mod/div operations to be replaced by masking. We assume sensible padding and alignment to reduce the impact of false sharing. (On x86 I recommend 128-byte alignment and padding because of the adjacent-sector prefetch facility). Each queue also has PutCursor and TakeCursor cursor variables, each of which should be sequestered as the sole occupant of a cache line or sector. You can opt to use 64-bit integers if concerned about wrap-around aliasing in the cursor variables. Put(null) is considered illegal, but the caller or implementation can easily check for and convert null to a distinguished non-null proxy value if null happens to be a value you'd like to pass. Take() will accordingly convert the proxy value back to null. An advantage of PTLQueue is that you can use atomic fetch-and-increment for the partial methods. We initialize each slot at index I with (Turn=I, MailBox=null). Both cursors are initially 0. All shared variables are considered "volatile" and atomics such as CAS and AtomicFetchAndIncrement are presumed to have bidirectional fence semantics. Finally T is the templated type. I've sketched out a total tryTake() method below that allows the caller to poll the queue. tryPut() has an analogous construction. Zebra stripping : alternating row colors for nice-looking code listings. See also google code "prettify" : https://code.google.com/p/google-code-prettify/ Prettify is a javascript module that yields the HTML/CSS/JS equivalent of pretty-print. -- pre:nth-child(odd) { background-color:#ff0000; } pre:nth-child(even) { background-color:#0000ff; } border-left: 11px solid #ccc; margin: 1.7em 0 1.7em 0.3em; background-color:#BFB; font-size:12px; line-height:65%; " // PTLQueue : Put(v) : // producer : partial method - waits as necessary assert v != null assert Mask = 1 && (Mask & (Mask+1)) == 0 // Document invariants // doorway step // Obtain a sequence number -- ticket // As a practical concern the ticket value is temporally unique // The ticket also identifies and selects a slot auto tkt = AtomicFetchIncrement (&PutCursor, 1) slot * s = &Slots[tkt & Mask] // waiting phase : // wait for slot's generation to match the tkt value assigned to this put() invocation. // The "generation" is implicitly encoded as the upper bits in the cursor // above those used to specify the index : tkt div (Mask+1) // The generation serves as an epoch number to identify a cohort of threads // accessing disjoint slots while s-Turn != tkt : Pause assert s-MailBox == null s-MailBox = v // deposit and pass message Take() : // consumer : partial method - waits as necessary auto tkt = AtomicFetchIncrement (&TakeCursor,1) slot * s = &Slots[tkt & Mask] // 2-stage waiting : // First wait for turn for our generation // Acquire exclusive "take" access to slot's MailBox field // Then wait for the slot to become occupied while s-Turn != tkt : Pause // Concurrency in this section of code is now reduced to just 1 producer thread // vs 1 consumer thread. // For a given queue and slot, there will be most one Take() operation running // in this section. // Consumer waits for producer to arrive and make slot non-empty // Extract message; clear mailbox; advance Turn indicator // We have an obvious happens-before relation : // Put(m) happens-before corresponding Take() that returns that same "m" for T v = s-MailBox if v != null : s-MailBox = null ST-ST barrier s-Turn = tkt + Mask + 1 // unlock slot to admit next producer and consumer return v Pause tryTake() : // total method - returns ASAP with failure indication for auto tkt = TakeCursor slot * s = &Slots[tkt & Mask] if s-Turn != tkt : return null T v = s-MailBox // presumptive return value if v == null : return null // ratify tkt and v values and commit by advancing cursor if CAS (&TakeCursor, tkt, tkt+1) != tkt : continue s-MailBox = null ST-ST barrier s-Turn = tkt + Mask + 1 return v The basic idea derives from the Partitioned Ticket Lock "PTL" (US20120240126-A1) and the MultiLane Concurrent Bag (US8689237). The latter is essentially a circular ring-buffer where the elements themselves are queues or concurrent collections. You can think of the PTLQueue as a partitioned ticket lock "PTL" augmented to pass values from lock to unlock via the slots. Alternatively, you could conceptualize of PTLQueue as a degenerate MultiLane bag where each slot or "lane" consists of a simple single-word MailBox instead of a general queue. Each lane in PTLQueue also has a private Turn field which acts like the Turn (Grant) variables found in PTL. Turn enforces strict FIFO ordering and restricts concurrency on the slot mailbox field to at most one simultaneous put() and take() operation. PTL uses a single "ticket" variable and per-slot Turn (grant) fields while MultiLane has distinct PutCursor and TakeCursor cursors and abstract per-slot sub-queues. Both PTL and MultiLane advance their cursor and ticket variables with atomic fetch-and-increment. PTLQueue borrows from both PTL and MultiLane and has distinct put and take cursors and per-slot Turn fields. Instead of a per-slot queues, PTLQueue uses a simple single-word MailBox field. PutCursor and TakeCursor act like a pair of ticket locks, conferring "put" and "take" access to a given slot. PutCursor, for instance, assigns an incoming put() request to a slot and serves as a PTL "Ticket" to acquire "put" permission to that slot's MailBox field. To better explain the operation of PTLQueue we deconstruct the operation of put() and take() as follows. Put() first increments PutCursor obtaining a new unique ticket. That ticket value also identifies a slot. Put() next waits for that slot's Turn field to match that ticket value. This is tantamount to using a PTL to acquire "put" permission on the slot's MailBox field. Finally, having obtained exclusive "put" permission on the slot, put() stores the message value into the slot's MailBox. Take() similarly advances TakeCursor, identifying a slot, and then acquires and secures "take" permission on a slot by waiting for Turn. Take() then waits for the slot's MailBox to become non-empty, extracts the message, and clears MailBox. Finally, take() advances the slot's Turn field, which releases both "put" and "take" access to the slot's MailBox. Note the asymmetry : put() acquires "put" access to the slot, but take() releases that lock. At any given time, for a given slot in a PTLQueue, at most one thread has "put" access and at most one thread has "take" access. This restricts concurrency from general MPMC to 1-vs-1. We have 2 ticket locks -- one for put() and one for take() -- each with its own "ticket" variable in the form of the corresponding cursor, but they share a single "Grant" egress variable in the form of the slot's Turn variable. Advancing the PutCursor, for instance, serves two purposes. First, we obtain a unique ticket which identifies a slot. Second, incrementing the cursor is the doorway protocol step to acquire the per-slot mutual exclusion "put" lock. The cursors and operations to increment those cursors serve double-duty : slot-selection and ticket assignment for locking the slot's MailBox field. At any given time a slot MailBox field can be in one of the following states: empty with no pending operations -- neutral state; empty with one or more waiting take() operations pending -- deficit; occupied with no pending operations; occupied with one or more waiting put() operations -- surplus; empty with a pending put() or pending put() and take() operations -- transitional; or occupied with a pending take() or pending put() and take() operations -- transitional. The partial put() and take() operators can be implemented with an atomic fetch-and-increment operation, which may confer a performance advantage over a CAS-based loop. In addition we have independent PutCursor and TakeCursor cursors. Critically, a put() operation modifies PutCursor but does not access the TakeCursor and a take() operation modifies the TakeCursor cursor but does not access the PutCursor. This acts to reduce coherence traffic relative to some other queue designs. It's worth noting that slow threads or obstruction in one slot (or "lane") does not impede or obstruct operations in other slots -- this gives us some degree of obstruction isolation. PTLQueue is not lock-free, however. The implementation above is expressed with polite busy-waiting (Pause) but it's trivial to implement per-slot parking and unparking to deschedule waiting threads. It's also easy to convert the queue to a more general deque by replacing the PutCursor and TakeCursor cursors with Left/Front and Right/Back cursors that can move either direction. Specifically, to push and pop from the "left" side of the deque we would decrement and increment the Left cursor, respectively, and to push and pop from the "right" side of the deque we would increment and decrement the Right cursor, respectively. We used a variation of PTLQueue for message passing in our recent OPODIS 2013 paper. ul { list-style:none; padding-left:0; padding:0; margin:0; margin-left:0; } ul#myTagID { padding: 0px; margin: 0px; list-style:none; margin-left:0;} -- -- There's quite a bit of related literature in this area. I'll call out a few relevant references: Wilson's NYU Courant Institute UltraComputer dissertation from 1988 is classic and the canonical starting point : Operating System Data Structures for Shared-Memory MIMD Machines with Fetch-and-Add. Regarding provenance and priority, I think PTLQueue or queues effectively equivalent to PTLQueue have been independently rediscovered a number of times. See CB-Queue and BNPBV, below, for instance. But Wilson's dissertation anticipates the basic idea and seems to predate all the others. Gottlieb et al : Basic Techniques for the Efficient Coordination of Very Large Numbers of Cooperating Sequential Processors Orozco et al : CB-Queue in Toward high-throughput algorithms on many-core architectures which appeared in TACO 2012. Meneghin et al : BNPVB family in Performance evaluation of inter-thread communication mechanisms on multicore/multithreaded architecture Dmitry Vyukov : bounded MPMC queue (highly recommended) Alex Otenko : US8607249 (highly related). John Mellor-Crummey : Concurrent queues: Practical fetch-and-phi algorithms. Technical Report 229, Department of Computer Science, University of Rochester Thomasson : FIFO Distributed Bakery Algorithm (very similar to PTLQueue). Scott and Scherer : Dual Data Structures I'll propose an optimization left as an exercise for the reader. Say we wanted to reduce memory usage by eliminating inter-slot padding. Such padding is usually "dark" memory and otherwise unused and wasted. But eliminating the padding leaves us at risk of increased false sharing. Furthermore lets say it was usually the case that the PutCursor and TakeCursor were numerically close to each other. (That's true in some use cases). We might still reduce false sharing by incrementing the cursors by some value other than 1 that is not trivially small and is coprime with the number of slots. Alternatively, we might increment the cursor by one and mask as usual, resulting in a logical index. We then use that logical index value to index into a permutation table, yielding an effective index for use in the slot array. The permutation table would be constructed so that nearby logical indices would map to more distant effective indices. (Open question: what should that permutation look like? Possibly some perversion of a Gray code or De Bruijn sequence might be suitable). As an aside, say we need to busy-wait for some condition as follows : "while C == 0 : Pause". Lets say that C is usually non-zero, so we typically don't wait. But when C happens to be 0 we'll have to spin for some period, possibly brief. We can arrange for the code to be more machine-friendly with respect to the branch predictors by transforming the loop into : "if C == 0 : for { Pause; if C != 0 : break; }". Critically, we want to restructure the loop so there's one branch that controls entry and another that controls loop exit. A concern is that your compiler or JIT might be clever enough to transform this back to "while C == 0 : Pause". You can sometimes avoid this by inserting a call to a some type of very cheap "opaque" method that the compiler can't elide or reorder. On Solaris, for instance, you could use :"if C == 0 : { gethrtime(); for { Pause; if C != 0 : break; }}". It's worth noting the obvious duality between locks and queues. If you have strict FIFO lock implementation with local spinning and succession by direct handoff such as MCS or CLH,then you can usually transform that lock into a queue. Hidden commentary and annotations - invisible : * And of course there's a well-known duality between queues and locks, but I'll leave that topic for another blog post. * Compare and contrast : PTLQ vs PTL and MultiLane * Equivalent : Turn; seq; sequence; pos; position; ticket * Put = Lock; Deposit Take = identify and reserve slot; wait; extract & clear; unlock * conceptualize : Distinct PutLock and TakeLock implemented as ticket lock or PTL Distinct arrival cursors but share per-slot "Turn" variable provides exclusive role-based access to slot's mailbox field put() acquires exclusive access to a slot for purposes of "deposit" assigns slot round-robin and then acquires deposit access rights/perms to that slot take() acquires exclusive access to slot for purposes of "withdrawal" assigns slot round-robin and then acquires withdrawal access rights/perms to that slot At any given time, only one thread can have withdrawal access to a slot at any given time, only one thread can have deposit access to a slot Permissible for T1 to have deposit access and T2 to simultaneously have withdrawal access * round-robin for the purposes of; role-based; access mode; access role mailslot; mailbox; allocate/assign/identify slot rights; permission; license; access permission; * PTL/Ticket hybrid Asymmetric usage ; owner oblivious lock-unlock pairing K-exclusion add Grant cursor pass message m from lock to unlock via Slots[] array Cursor performs 2 functions : + PTL ticket + Assigns request to slot in round-robin fashion Deconstruct protocol : explication put() : allocate slot in round-robin fashion acquire PTL for "put" access store message into slot associated with PTL index take() : Acquire PTL for "take" access // doorway step seq = fetchAdd (&Grant, 1) s = &Slots[seq & Mask] // waiting phase while s-Turn != seq : pause Extract : wait for s-mailbox to be full v = s-mailbox s-mailbox = null Release PTL for both "put" and "take" access s-Turn = seq + Mask + 1 * Slot round-robin assignment and lock "doorway" protocol leverage the same cursor and FetchAdd operation on that cursor FetchAdd (&Cursor,1) + round-robin slot assignment and dispersal + PTL/ticket lock "doorway" step waiting phase is via "Turn" field in slot * PTLQueue uses 2 cursors -- put and take. Acquire "put" access to slot via PTL-like lock Acquire "take" access to slot via PTL-like lock 2 locks : put and take -- at most one thread can access slot's mailbox Both locks use same "turn" field Like multilane : 2 cursors : put and take slot is simple 1-capacity mailbox instead of queue Borrow per-slot turn/grant from PTL Provides strict FIFO Lock slot : put-vs-put take-vs-take at most one put accesses slot at any one time at most one put accesses take at any one time reduction to 1-vs-1 instead of N-vs-M concurrency Per slot locks for put/take Release put/take by advancing turn * is instrumental in ... * P-V Semaphore vs lock vs K-exclusion * See also : FastQueues-excerpt.java dice-etc/queue-mpmc-bounded-blocking-circular-xadd/ * PTLQueue is the same as PTLQB - identical * Expedient return; ASAP; prompt; immediately * Lamport's Bakery algorithm : doorway step then waiting phase Threads arriving at doorway obtain a unique ticket number Threads enter in ticket order * In the terminology of Reed and Kanodia a ticket lock corresponds to the busy-wait implementation of a semaphore using an eventcount and a sequencer It can also be thought of as an optimization of Lamport's bakery lock was designed for fault-tolerance rather than performance Instead of spinning on the release counter, processors using a bakery lock repeatedly examine the tickets of their peers --

    Read the article

  • Performance Enhancement in Full-Text Search Query

    - by Calvin Sun
    Ever since its first release, we are continuing consolidating and developing InnoDB Full-Text Search feature. There is one recent improvement that worth blogging about. It is an effort with MySQL Optimizer team that simplifies some common queries’ Query Plans and dramatically shorted the query time. I will describe the issue, our solution and the end result by some performance numbers to demonstrate our efforts in continuing enhancement the Full-Text Search capability. The Issue: As we had discussed in previous Blogs, InnoDB implements Full-Text index as reversed auxiliary tables. The query once parsed will be reinterpreted into several queries into related auxiliary tables and then results are merged and consolidated to come up with the final result. So at the end of the query, we’ll have all matching records on hand, sorted by their ranking or by their Doc IDs. Unfortunately, MySQL’s optimizer and query processing had been initially designed for MyISAM Full-Text index, and sometimes did not fully utilize the complete result package from InnoDB. Here are a couple examples: Case 1: Query result ordered by Rank with only top N results: mysql> SELECT FTS_DOC_ID, MATCH (title, body) AGAINST ('database') AS SCORE FROM articles ORDER BY score DESC LIMIT 1; In this query, user tries to retrieve a single record with highest ranking. It should have a quick answer once we have all the matching documents on hand, especially if there are ranked. However, before this change, MySQL would almost retrieve rankings for almost every row in the table, sort them and them come with the top rank result. This whole retrieve and sort is quite unnecessary given the InnoDB already have the answer. In a real life case, user could have millions of rows, so in the old scheme, it would retrieve millions of rows' ranking and sort them, even if our FTS already found there are two 3 matched rows. Apparently, the million ranking retrieve is done in vain. In above case, it should just ask for 3 matched rows' ranking, all other rows' ranking are 0. If it want the top ranking, then it can just get the first record from our already sorted result. Case 2: Select Count(*) on matching records: mysql> SELECT COUNT(*) FROM articles WHERE MATCH (title,body) AGAINST ('database' IN NATURAL LANGUAGE MODE); In this case, InnoDB search can find matching rows quickly and will have all matching rows. However, before our change, in the old scheme, every row in the table was requested by MySQL one by one, just to check whether its ranking is larger than 0, and later comes up a count. In fact, there is no need for MySQL to fetch all rows, instead InnoDB already had all the matching records. The only thing need is to call an InnoDB API to retrieve the count The difference can be huge. Following query output shows how big the difference can be: mysql> select count(*) from searchindex_inno where match(si_title, si_text) against ('people')  +----------+ | count(*) | +----------+ | 666877 | +----------+ 1 row in set (16 min 17.37 sec) So the query took almost 16 minutes. Let’s see how long the InnoDB can come up the result. In InnoDB, you can obtain extra diagnostic printout by turning on “innodb_ft_enable_diag_print”, this will print out extra query info: Error log: keynr=2, 'people' NL search Total docs: 10954826 Total words: 0 UNION: Searching: 'people' Processing time: 2 secs: row(s) 666877: error: 10 ft_init() ft_init_ext() keynr=2, 'people' NL search Total docs: 10954826 Total words: 0 UNION: Searching: 'people' Processing time: 3 secs: row(s) 666877: error: 10 Output shows it only took InnoDB only 3 seconds to get the result, while the whole query took 16 minutes to finish. So large amount of time has been wasted on the un-needed row fetching. The Solution: The solution is obvious. MySQL can skip some of its steps, optimize its plan and obtain useful information directly from InnoDB. Some of savings from doing this include: 1) Avoid redundant sorting. Since InnoDB already sorted the result according to ranking. MySQL Query Processing layer does not need to sort to get top matching results. 2) Avoid row by row fetching to get the matching count. InnoDB provides all the matching records. All those not in the result list should all have ranking of 0, and no need to be retrieved. And InnoDB has a count of total matching records on hand. No need to recount. 3) Covered index scan. InnoDB results always contains the matching records' Document ID and their ranking. So if only the Document ID and ranking is needed, there is no need to go to user table to fetch the record itself. 4) Narrow the search result early, reduce the user table access. If the user wants to get top N matching records, we do not need to fetch all matching records from user table. We should be able to first select TOP N matching DOC IDs, and then only fetch corresponding records with these Doc IDs. Performance Results and comparison with MyISAM The result by this change is very obvious. I includes six testing result performed by Alexander Rubin just to demonstrate how fast the InnoDB query now becomes when comparing MyISAM Full-Text Search. These tests are base on the English Wikipedia data of 5.4 Million rows and approximately 16G table. The test was performed on a machine with 1 CPU Dual Core, SSD drive, 8G of RAM and InnoDB_buffer_pool is set to 8 GB. Table 1: SELECT with LIMIT CLAUSE mysql> SELECT si_title, match(si_title, si_text) against('family') as rel FROM si WHERE match(si_title, si_text) against('family') ORDER BY rel desc LIMIT 10; InnoDB MyISAM Times Faster Time for the query 1.63 sec 3 min 26.31 sec 127 You can see for this particular query (retrieve top 10 records), InnoDB Full-Text Search is now approximately 127 times faster than MyISAM. Table 2: SELECT COUNT QUERY mysql>select count(*) from si where match(si_title, si_text) against('family‘); +----------+ | count(*) | +----------+ | 293955 | +----------+ InnoDB MyISAM Times Faster Time for the query 1.35 sec 28 min 59.59 sec 1289 In this particular case, where there are 293k matching results, InnoDB took only 1.35 second to get all of them, while take MyISAM almost half an hour, that is about 1289 times faster!. Table 3: SELECT ID with ORDER BY and LIMIT CLAUSE for selected terms mysql> SELECT <ID>, match(si_title, si_text) against(<TERM>) as rel FROM si_<TB> WHERE match(si_title, si_text) against (<TERM>) ORDER BY rel desc LIMIT 10; Term InnoDB (time to execute) MyISAM(time to execute) Times Faster family 0.5 sec 5.05 sec 10.1 family film 0.95 sec 25.39 sec 26.7 Pizza restaurant orange county California 0.93 sec 32.03 sec 34.4 President united states of America 2.5 sec 36.98 sec 14.8 Table 4: SELECT title and text with ORDER BY and LIMIT CLAUSE for selected terms mysql> SELECT <ID>, si_title, si_text, ... as rel FROM si_<TB> WHERE match(si_title, si_text) against (<TERM>) ORDER BY rel desc LIMIT 10; Term InnoDB (time to execute) MyISAM(time to execute) Times Faster family 0.61 sec 41.65 sec 68.3 family film 1.15 sec 47.17 sec 41.0 Pizza restaurant orange county california 1.03 sec 48.2 sec 46.8 President united states of america 2.49 sec 44.61 sec 17.9 Table 5: SELECT ID with ORDER BY and LIMIT CLAUSE for selected terms mysql> SELECT <ID>, match(si_title, si_text) against(<TERM>) as rel  FROM si_<TB> WHERE match(si_title, si_text) against (<TERM>) ORDER BY rel desc LIMIT 10; Term InnoDB (time to execute) MyISAM(time to execute) Times Faster family 0.5 sec 5.05 sec 10.1 family film 0.95 sec 25.39 sec 26.7 Pizza restaurant orange county califormia 0.93 sec 32.03 sec 34.4 President united states of america 2.5 sec 36.98 sec 14.8 Table 6: SELECT COUNT(*) mysql> SELECT count(*) FROM si_<TB> WHERE match(si_title, si_text) against (<TERM>) LIMIT 10; Term InnoDB (time to execute) MyISAM(time to execute) Times Faster family 0.47 sec 82 sec 174.5 family film 0.83 sec 131 sec 157.8 Pizza restaurant orange county califormia 0.74 sec 106 sec 143.2 President united states of america 1.96 sec 220 sec 112.2  Again, table 3 to table 6 all showing InnoDB consistently outperform MyISAM in these queries by a large margin. It becomes obvious the InnoDB has great advantage over MyISAM in handling large data search. Summary: These results demonstrate the great performance we could achieve by making MySQL optimizer and InnoDB Full-Text Search more tightly coupled. I think there are still many cases that InnoDB’s result info have not been fully taken advantage of, which means we still have great room to improve. And we will continuously explore the area, and get more dramatic results for InnoDB full-text searches. Jimmy Yang, September 29, 2012

    Read the article

  • July, the 31 Days of SQL Server DMO’s – Day 26 (sys.dm_db_log_space_usage)

    - by Tamarick Hill
    The sys.dm_db_log_space_usage DMV is a new DMV for SQL Server 2012. It returns Total Size, Used Size, and Used Percent size for a transaction log file of a given database. To illustrate this DMV, I will query the DMV against my AdventureWorks2012 database. SELECT * FROM sys.dm_db_log_space_usage As mentioned above, the result set gives us the total size of the transaction log in bytes, the used size of the log in bytes, and the percent of the log that has been used. This is a very simplistic DMV but returns valuable information. Being able to detect when a transaction log is close to being full is always a valuable thing to alert on, and this DMV just provided an additional method for acquiring the necessary information. Follow me on Twitter @PrimeTimeDBA

    Read the article

  • Measure load of ubuntu server

    - by user84471
    I have Ubuntu server and apache2 installed on it. I want to benchmark it using ab tool from another machine. I want to know how measure load of this program. In top is showing this: top - 06:02:19 up 3 min, 2 users, load average: 0.28, 0.26, 0.12 Tasks: 94 total, 2 running, 92 sleeping, 0 stopped, 0 zombie Cpu(s): 67.2%us, 5.5%sy, 0.0%ni, 26.3%id, 0.5%wa, 0.0%hi, 0.5%si, 0.0%st Mem: 499320k total, 342028k used, 157292k free, 42504k buffers Swap: 514044k total, 0k used, 514044k free, 71388k cached PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND 1521 www-data 20 0 49180 22m 4144 S 23 4.6 0:00.68 apache2 1520 www-data 20 0 49180 22m 4152 S 22 4.6 0:00.66 apache2 1124 www-data 20 0 49180 22m 4152 S 22 4.6 0:00.70 apache2 1519 www-data 20 0 49180 22m 4152 S 20 4.6 0:00.70 apache2 1525 www-data 20 0 49180 22m 4144 S 16 4.6 0:00.47 apache2 1123 www-data 20 0 49168 22m 4152 S 9 4.6 0:00.78 apache2 1125 www-data 20 0 49168 22m 4152 S 9 4.6 0:00.77 apache2 1126 www-data 20 0 49168 22m 4152 S 9 4.6 0:00.73 apache2 1127 www-data 20 0 49168 22m 4152 S 9 4.6 0:00.74 apache2 1523 www-data 20 0 42896 14m 3388 R 5 3.1 0:00.14 apache2 1038 whoopsie 20 0 24476 3720 2856 S 0 0.7 0:00.01 whoopsie 1089 root 20 0 34060 6988 3640 S 0 1.4 0:00.05 apache2 1338 ubuntu 20 0 2832 1192 944 S 0 0.2 0:00.30 top 1417 ubuntu 20 0 9652 1456 824 S 0 0.3 0:00.02 sshd 1 root 20 0 3540 1876 1248 S 0 0.4 0:00.83 init 2 root 20 0 0 0 0 S 0 0.0 0:00.00 kthreadd 3 root 20 0 0 0 0 S 0 0.0 0:00.03 ksoftirqd/0 4 root 20 0 0 0 0 S 0 0.0 0:00.00 kworker/0:0 5 root 20 0 0 0 0 S 0 0.0 0:00.48 kworker/u:0 6 root RT 0 0 0 0 S 0 0.0 0:00.00 migration/0 7 root RT 0 0 0 0 S 0 0.0 0:00.00 watchdog/0 8 root RT 0 0 0 0 S 0 0.0 0:00.00 migration/1 9 root 20 0 0 0 0 S 0 0.0 0:00.00 kworker/1:0 10 root 20 0 0 0 0 S 0 0.0 0:00.01 ksoftirqd/1 11 root 20 0 0 0 0 S 0 0.0 0:00.17 kworker/0:1 12 root RT 0 0 0 0 S 0 0.0 0:00.00 watchdog/1 13 root 0 -20 0 0 0 S 0 0.0 0:00.00 cpuset 14 root 0 -20 0 0 0 S 0 0.0 0:00.00 khelper 15 root 20 0 0 0 0 S 0 0.0 0:00.00 kdevtmpfs 16 root 0 -20 0 0 0 S 0 0.0 0:00.00 netns 17 root 20 0 0 0 0 S 0 0.0 0:00.32 kworker/u:1 18 root 20 0 0 0 0 S 0 0.0 0:00.00 sync_supers 19 root 20 0 0 0 0 S 0 0.0 0:00.00 bdi-default 20 root 0 -20 0 0 0 S 0 0.0 0:00.00 kintegrityd 21 root 0 -20 0 0 0 S 0 0.0 0:00.00 kblockd 22 root 0 -20 0 0 0 S 0 0.0 0:00.00 ata_sff 23 root 20 0 0 0 0 S 0 0.0 0:00.00 khubd So its difficult to do it without any script or program. How can I do this?

    Read the article

< Previous Page | 29 30 31 32 33 34 35 36 37 38 39 40  | Next Page >