Search Results

Search found 8692 results on 348 pages for 'per magnusson'.

Page 51/348 | < Previous Page | 47 48 49 50 51 52 53 54 55 56 57 58  | Next Page >

  • Light following me around the room. Something is wrong with my shader!

    - by Robinson
    I'm trying to do a spot (Blinn) light, with falloff and attenuation. It seems to be working OK except I have a bit of a space problem. That is, whenever I move the camera the light moves to maintain the same relative position, rather than changing with the camera. This results in the light moving around, i.e. not always falling on the same surfaces. It's as if there's a flashlight attached to the camera. I'm transforming the lights beforehand into view space, so Light_Position and Light_Direction are already in eye space (I hope!). I made a little movie of what it looks like here: My camera rotating around a point inside a box. The light is fixed in the centre up and its "look at" point in a fixed position in front of it. As you can see, as the camera rotates around the origin (always looking at the centre), so don't think the box is rotating (!). The lighting follows it around. To start, some code. This is how I'm transforming the light into view space (it gets passed into the shader already in view space): // Compute eye-space light position. Math::Vector3d eyeSpacePosition = MyCamera->ViewMatrix() * MyLightPosition; MyShaderVariables->Set(MyLightPositionIndex, eyeSpacePosition); // Compute eye-space light direction vector. Math::Vector3d eyeSpaceDirection = Math::Unit(MyLightLookAt - MyLightPosition); MyCamera->ViewMatrixInverseTranspose().TransformNormal(eyeSpaceDirection); MyShaderVariables->Set(MyLightDirectionIndex, eyeSpaceDirection); Can anyone give me a clue as to what I'm doing wrong here? I think the light should remain looking at a fixed point on the box, regardless of the camera orientation. Here are the vertex and pixel shaders: /////////////////////////////////////////////////// // Vertex Shader /////////////////////////////////////////////////// #version 420 /////////////////////////////////////////////////// // Uniform Buffer Structures /////////////////////////////////////////////////// // Camera. layout (std140) uniform Camera { mat4 Camera_View; mat4 Camera_ViewInverseTranspose; mat4 Camera_Projection; }; // Matrices per model. layout (std140) uniform Model { mat4 Model_World; mat4 Model_WorldView; mat4 Model_WorldViewInverseTranspose; mat4 Model_WorldViewProjection; }; // Spotlight. layout (std140) uniform OmniLight { float Light_Intensity; vec3 Light_Position; vec3 Light_Direction; vec4 Light_Ambient_Colour; vec4 Light_Diffuse_Colour; vec4 Light_Specular_Colour; float Light_Attenuation_Min; float Light_Attenuation_Max; float Light_Cone_Min; float Light_Cone_Max; }; /////////////////////////////////////////////////// // Streams (per vertex) /////////////////////////////////////////////////// layout(location = 0) in vec3 attrib_Position; layout(location = 1) in vec3 attrib_Normal; layout(location = 2) in vec3 attrib_Tangent; layout(location = 3) in vec3 attrib_BiNormal; layout(location = 4) in vec2 attrib_Texture; /////////////////////////////////////////////////// // Output streams (per vertex) /////////////////////////////////////////////////// out vec3 attrib_Fragment_Normal; out vec4 attrib_Fragment_Position; out vec2 attrib_Fragment_Texture; out vec3 attrib_Fragment_Light; out vec3 attrib_Fragment_Eye; /////////////////////////////////////////////////// // Main /////////////////////////////////////////////////// void main() { // Transform normal into eye space attrib_Fragment_Normal = (Model_WorldViewInverseTranspose * vec4(attrib_Normal, 0.0)).xyz; // Transform vertex into eye space (world * view * vertex = eye) vec4 position = Model_WorldView * vec4(attrib_Position, 1.0); // Compute vector from eye space vertex to light (light is in eye space already) attrib_Fragment_Light = Light_Position - position.xyz; // Compute vector from the vertex to the eye (which is now at the origin). attrib_Fragment_Eye = -position.xyz; // Output texture coord. attrib_Fragment_Texture = attrib_Texture; // Compute vertex position by applying camera projection. gl_Position = Camera_Projection * position; } and the pixel shader: /////////////////////////////////////////////////// // Pixel Shader /////////////////////////////////////////////////// #version 420 /////////////////////////////////////////////////// // Samplers /////////////////////////////////////////////////// uniform sampler2D Map_Diffuse; /////////////////////////////////////////////////// // Global Uniforms /////////////////////////////////////////////////// // Material. layout (std140) uniform Material { vec4 Material_Ambient_Colour; vec4 Material_Diffuse_Colour; vec4 Material_Specular_Colour; vec4 Material_Emissive_Colour; float Material_Shininess; float Material_Strength; }; // Spotlight. layout (std140) uniform OmniLight { float Light_Intensity; vec3 Light_Position; vec3 Light_Direction; vec4 Light_Ambient_Colour; vec4 Light_Diffuse_Colour; vec4 Light_Specular_Colour; float Light_Attenuation_Min; float Light_Attenuation_Max; float Light_Cone_Min; float Light_Cone_Max; }; /////////////////////////////////////////////////// // Input streams (per vertex) /////////////////////////////////////////////////// in vec3 attrib_Fragment_Normal; in vec3 attrib_Fragment_Position; in vec2 attrib_Fragment_Texture; in vec3 attrib_Fragment_Light; in vec3 attrib_Fragment_Eye; /////////////////////////////////////////////////// // Result /////////////////////////////////////////////////// out vec4 Out_Colour; /////////////////////////////////////////////////// // Main /////////////////////////////////////////////////// void main(void) { // Compute N dot L. vec3 N = normalize(attrib_Fragment_Normal); vec3 L = normalize(attrib_Fragment_Light); vec3 E = normalize(attrib_Fragment_Eye); vec3 H = normalize(L + E); float NdotL = clamp(dot(L,N), 0.0, 1.0); float NdotH = clamp(dot(N,H), 0.0, 1.0); // Compute ambient term. vec4 ambient = Material_Ambient_Colour * Light_Ambient_Colour; // Diffuse. vec4 diffuse = texture2D(Map_Diffuse, attrib_Fragment_Texture) * Light_Diffuse_Colour * Material_Diffuse_Colour * NdotL; // Specular. float specularIntensity = pow(NdotH, Material_Shininess) * Material_Strength; vec4 specular = Light_Specular_Colour * Material_Specular_Colour * specularIntensity; // Light attenuation (so we don't have to use 1 - x, we step between Max and Min). float d = length(-attrib_Fragment_Light); float attenuation = smoothstep(Light_Attenuation_Max, Light_Attenuation_Min, d); // Adjust attenuation based on light cone. float LdotS = dot(-L, Light_Direction), CosI = Light_Cone_Min - Light_Cone_Max; attenuation *= clamp((LdotS - Light_Cone_Max) / CosI, 0.0, 1.0); // Final colour. Out_Colour = (ambient + diffuse + specular) * Light_Intensity * attenuation; }

    Read the article

  • Organization &amp; Architecture UNISA Studies &ndash; Chap 4

    - by MarkPearl
    Learning Outcomes Explain the characteristics of memory systems Describe the memory hierarchy Discuss cache memory principles Discuss issues relevant to cache design Describe the cache organization of the Pentium Computer Memory Systems There are key characteristics of memory… Location – internal or external Capacity – expressed in terms of bytes Unit of Transfer – the number of bits read out of or written into memory at a time Access Method – sequential, direct, random or associative From a users perspective the two most important characteristics of memory are… Capacity Performance – access time, memory cycle time, transfer rate The trade off for memory happens along three axis… Faster access time, greater cost per bit Greater capacity, smaller cost per bit Greater capacity, slower access time This leads to people using a tiered approach in their use of memory   As one goes down the hierarchy, the following occurs… Decreasing cost per bit Increasing capacity Increasing access time Decreasing frequency of access of the memory by the processor The use of two levels of memory to reduce average access time works in principle, but only if conditions 1 to 4 apply. A variety of technologies exist that allow us to accomplish this. Thus it is possible to organize data across the hierarchy such that the percentage of accesses to each successively lower level is substantially less than that of the level above. A portion of main memory can be used as a buffer to hold data temporarily that is to be read out to disk. This is sometimes referred to as a disk cache and improves performance in two ways… Disk writes are clustered. Instead of many small transfers of data, we have a few large transfers of data. This improves disk performance and minimizes processor involvement. Some data designed for write-out may be referenced by a program before the next dump to disk. In that case the data is retrieved rapidly from the software cache rather than slowly from disk. Cache Memory Principles Cache memory is substantially faster than main memory. A caching system works as follows.. When a processor attempts to read a word of memory, a check is made to see if this in in cache memory… If it is, the data is supplied, If it is not in the cache, a block of main memory, consisting of a fixed number of words is loaded to the cache. Because of the phenomenon of locality of references, when a block of data is fetched into the cache, it is likely that there will be future references to that same memory location or to other words in the block. Elements of Cache Design While there are a large number of cache implementations, there are a few basic design elements that serve to classify and differentiate cache architectures… Cache Addresses Cache Size Mapping Function Replacement Algorithm Write Policy Line Size Number of Caches Cache Addresses Almost all non-embedded processors support virtual memory. Virtual memory in essence allows a program to address memory from a logical point of view without needing to worry about the amount of physical memory available. When virtual addresses are used the designer may choose to place the cache between the MMU (memory management unit) and the processor or between the MMU and main memory. The disadvantage of virtual memory is that most virtual memory systems supply each application with the same virtual memory address space (each application sees virtual memory starting at memory address 0), which means the cache memory must be completely flushed with each application context switch or extra bits must be added to each line of the cache to identify which virtual address space the address refers to. Cache Size We would like the size of the cache to be small enough so that the overall average cost per bit is close to that of main memory alone and large enough so that the overall average access time is close to that of the cache alone. Also, larger caches are slightly slower than smaller ones. Mapping Function Because there are fewer cache lines than main memory blocks, an algorithm is needed for mapping main memory blocks into cache lines. The choice of mapping function dictates how the cache is organized. Three techniques can be used… Direct – simplest technique, maps each block of main memory into only one possible cache line Associative – Each main memory block to be loaded into any line of the cache Set Associative – exhibits the strengths of both the direct and associative approaches while reducing their disadvantages For detailed explanations of each approach – read the text book (page 148 – 154) Replacement Algorithm For associative and set associating mapping a replacement algorithm is needed to determine which of the existing blocks in the cache must be replaced by a new block. There are four common approaches… LRU (Least recently used) FIFO (First in first out) LFU (Least frequently used) Random selection Write Policy When a block resident in the cache is to be replaced, there are two cases to consider If no writes to that block have happened in the cache – discard it If a write has occurred, a process needs to be initiated where the changes in the cache are propagated back to the main memory. There are several approaches to achieve this including… Write Through – all writes to the cache are done to the main memory as well at the point of the change Write Back – when a block is replaced, all dirty bits are written back to main memory The problem is complicated when we have multiple caches, there are techniques to accommodate for this but I have not summarized them. Line Size When a block of data is retrieved and placed in the cache, not only the desired word but also some number of adjacent words are retrieved. As the block size increases from very small to larger sizes, the hit ratio will at first increase because of the principle of locality, which states that the data in the vicinity of a referenced word are likely to be referenced in the near future. As the block size increases, more useful data are brought into cache. The hit ratio will begin to decrease as the block becomes even bigger and the probability of using the newly fetched information becomes less than the probability of using the newly fetched information that has to be replaced. Two specific effects come into play… Larger blocks reduce the number of blocks that fit into a cache. Because each block fetch overwrites older cache contents, a small number of blocks results in data being overwritten shortly after they are fetched. As a block becomes larger, each additional word is farther from the requested word and therefore less likely to be needed in the near future. The relationship between block size and hit ratio is complex, and no set approach is judged to be the best in all circumstances.   Pentium 4 and ARM cache organizations The processor core consists of four major components: Fetch/decode unit – fetches program instruction in order from the L2 cache, decodes these into a series of micro-operations, and stores the results in the L2 instruction cache Out-of-order execution logic – Schedules execution of the micro-operations subject to data dependencies and resource availability – thus micro-operations may be scheduled for execution in a different order than they were fetched from the instruction stream. As time permits, this unit schedules speculative execution of micro-operations that may be required in the future Execution units – These units execute micro-operations, fetching the required data from the L1 data cache and temporarily storing results in registers Memory subsystem – This unit includes the L2 and L3 caches and the system bus, which is used to access main memory when the L1 and L2 caches have a cache miss and to access the system I/O resources

    Read the article

  • Inequality joins, Asynchronous transformations and Lookups : SSIS

    - by jamiet
    It is pretty much accepted by SQL Server Integration Services (SSIS) developers that synchronous transformations are generally quicker than asynchronous transformations (for a description of synchronous and asynchronous transformations go read Asynchronous and synchronous data flow components). Notice I said “generally” and not “always”; there are circumstances where using asynchronous transformations can be beneficial and in this blog post I’ll demonstrate such a scenario, one that is pretty common when building data warehouses. Imagine I have a [Customer] dimension table that manages information about all of my customers as a slowly-changing dimension. If that is a type 2 slowly changing dimension then you will likely have multiple rows per customer in that table. Furthermore you might also have datetime fields that indicate the effective time period of each member record. Here is such a table that contains data for four dimension members {Terry, Max, Henry, Horace}: Notice that we have multiple records per customer and that the [SCDStartDate] of a record is equivalent to the [SCDEndDate] of the record that preceded it (if there was one). (Note that I am on record as saying I am not a fan of this technique of storing an [SCDEndDate] but for the purposes of clarity I have included it here.) Anyway, the idea here is that we will have some incoming data containing [CustomerName] & [EffectiveDate] and we need to use those values to lookup [Customer].[CustomerId]. The logic will be: Lookup a [CustomerId] WHERE [CustomerName]=[CustomerName] AND [SCDStartDate] <= [EffectiveDate] AND [EffectiveDate] <= [SCDEndDate] The conventional approach to this would be to use a full cached lookup but that isn’t an option here because we are using inequality conditions. The obvious next step then is to use a non-cached lookup which enables us to change the SQL statement to use inequality operators: Let’s take a look at the dataflow: Notice these are all synchronous components. This approach works just fine however it does have the limitation that it has to issue a SQL statement against your lookup set for every row thus we can expect the execution time of our dataflow to increase linearly in line with the number of rows in our dataflow; that’s not good. OK, that’s the obvious method. Let’s now look at a different way of achieving this using an asynchronous Merge Join transform coupled with a Conditional Split. I’ve shown it post-execution so that I can include the row counts which help to illustrate what is going on here: Notice that there are more rows output from our Merge Join component than on the input. That is because we are joining on [CustomerName] and, as we know, we have multiple records per [CustomerName] in our lookup set. Notice also that there are two asynchronous components in here (the Sort and the Merge Join). I have embedded a video below that compares the execution times for each of these two methods. The video is just over 8minutes long. View on Vimeo  For those that can’t be bothered watching the video I’ll tell you the results here. The dataflow that used the Lookup transform took 36 seconds whereas the dataflow that used the Merge Join took less than two seconds. An illustration in case it is needed: Pretty conclusive proof that in some scenarios it may be quicker to use an asynchronous component than a synchronous one. Your mileage may of course vary. The scenario outlined here is analogous to performance tuning procedural SQL that uses cursors. It is common to eliminate cursors by converting them to set-based operations and that is effectively what we have done here. Our non-cached lookup is performing a discrete operation for every single row of data, exactly like a cursor does. By eliminating this cursor-in-disguise we have dramatically sped up our dataflow. I hope all of that proves useful. You can download the package that I demonstrated in the video from my SkyDrive at http://cid-550f681dad532637.skydrive.live.com/self.aspx/Public/BlogShare/20100514/20100514%20Lookups%20and%20Merge%20Joins.zip Comments are welcome as always. @Jamiet Share this post: email it! | bookmark it! | digg it! | reddit! | kick it! | live it!

    Read the article

  • Sun Fire X4800 M2 Posts World Record x86 SPECjEnterprise2010 Result

    - by Brian
    Oracle's Sun Fire X4800 M2 using the Intel Xeon E7-8870 processor and Sun Fire X4470 M2 using the Intel Xeon E7-4870 processor, produced a world record single application server SPECjEnterprise2010 benchmark result of 27,150.05 SPECjEnterprise2010 EjOPS. The Sun Fire X4800 M2 server ran the application tier and the Sun Fire X4470 M2 server was used for the database tier. The Sun Fire X4800 M2 server demonstrated 63% better performance compared to IBM P780 server result of 16,646.34 SPECjEnterprise2010 EjOPS. The Sun Fire X4800 M2 server demonstrated 4% better performance than the Cisco UCS B440 M2 result, both results used the same number of processors. This result used Oracle WebLogic Server 12c, Java HotSpot(TM) 64-Bit Server 1.7.0_02, and Oracle Database 11g. This result was produced using Oracle Linux. Performance Landscape Complete benchmark results are at the SPEC website, SPECjEnterprise2010 Results. The table below compares against the best results from IBM and Cisco. SPECjEnterprise2010 Performance Chart as of 3/12/2012 Submitter EjOPS* Application Server Database Server Oracle 27,150.05 1x Sun Fire X4800 M2 8x 2.4 GHz Intel Xeon E7-8870 Oracle WebLogic 12c 1x Sun Fire X4470 M2 4x 2.4 GHz Intel Xeon E7-4870 Oracle Database 11g (11.2.0.2) Cisco 26,118.67 2x UCS B440 M2 Blade Server 4x 2.4 GHz Intel Xeon E7-4870 Oracle WebLogic 11g (10.3.5) 1x UCS C460 M2 Blade Server 4x 2.4 GHz Intel Xeon E7-4870 Oracle Database 11g (11.2.0.2) IBM 16,646.34 1x IBM Power 780 8x 3.86 GHz POWER 7 WebSphere Application Server V7 1x IBM Power 750 Express 4x 3.55 GHz POWER 7 IBM DB2 9.7 Workgroup Server Edition FP3a * SPECjEnterprise2010 EjOPS, bigger is better. Configuration Summary Application Server: 1 x Sun Fire X4800 M2 8 x 2.4 GHz Intel Xeon processor E7-8870 256 GB memory 4 x 10 GbE NIC 2 x FC HBA Oracle Linux 5 Update 6 Oracle WebLogic Server 11g Release 1 (10.3.5) Java HotSpot(TM) 64-Bit Server VM on Linux, version 1.7.0_02 (Java SE 7 Update 2) Database Server: 1 x Sun Fire X4470 M2 4 x 2.4 GHz Intel Xeon E7-4870 512 GB memory 4 x 10 GbE NIC 2 x FC HBA 2 x Sun StorageTek 2540 M2 4 x Sun Fire X4270 M2 4 x Sun Storage F5100 Flash Array Oracle Linux 5 Update 6 Oracle Database 11g Enterprise Edition Release 11.2.0.2 Benchmark Description SPECjEnterprise2010 is the third generation of the SPEC organization's J2EE end-to-end industry standard benchmark application. The SPECjEnterprise2010 benchmark has been designed and developed to cover the Java EE 5 specification's significantly expanded and simplified programming model, highlighting the major features used by developers in the industry today. This provides a real world workload driving the Application Server's implementation of the Java EE specification to its maximum potential and allowing maximum stressing of the underlying hardware and software systems. The workload consists of an end to end web based order processing domain, an RMI and Web Services driven manufacturing domain and a supply chain model utilizing document based Web Services. The application is a collection of Java classes, Java Servlets, Java Server Pages, Enterprise Java Beans, Java Persistence Entities (pojo's) and Message Driven Beans. The SPECjEnterprise2010 benchmark heavily exercises all parts of the underlying infrastructure that make up the application environment, including hardware, JVM software, database software, JDBC drivers, and the system network. The primary metric of the SPECjEnterprise2010 benchmark is jEnterprise Operations Per Second ("SPECjEnterprise2010 EjOPS"). This metric is calculated by adding the metrics of the Dealership Management Application in the Dealer Domain and the Manufacturing Application in the Manufacturing Domain. There is no price/performance metric in this benchmark. Key Points and Best Practices Sixteen Oracle WebLogic server instances were started using numactl, binding 2 instances per chip. Eight Oracle database listener processes were started, binding 2 instances per chip using taskset. Additional tuning information is in the report at http://spec.org. See Also Oracle Press Release -- SPECjEnterprise2010 Results Page Sun Fire X4800 M2 Server oracle.com OTN Sun Fire X4270 M2 Server oracle.com OTN Sun Storage 2540-M2 Array oracle.com OTN Oracle Linux oracle.com OTN Oracle Database 11g Release 2 Enterprise Edition oracle.com OTN WebLogic Suite oracle.com OTN Disclosure Statement SPEC and the benchmark name SPECjEnterprise are registered trademarks of the Standard Performance Evaluation Corporation. Sun Fire X4800 M2, 27,150.05 SPECjEnterprise2010 EjOPS; IBM Power 780, 16,646.34 SPECjEnterprise2010 EjOPS; Cisco UCS B440 M2, 26,118.67 SPECjEnterprise2010 EjOPS. Results from www.spec.org as of 3/27/2012.

    Read the article

  • P90X or How I Stopped Worrying and Love Exercise

    - by Matt Christian
    Last Wednesday, after many UPS delivery failures, I received P90X in the mail.  P90X is a series of DVD's and a nutrition guide you use to shed pounds and gain muscle.  Odds are you've seen the infomercial on TV at some point if you watch a little tube now and again.  I started last Thursday and am still standing to tell this tale. At it's core, P90X is a 12 DVD set of exercise videos.  Each video is comprised of a different workout routine that typically last around an hour (some up to 1 1/2 hours).  Every day you are supposed to do one of the workouts which are different every day (sometimes you may repeat a shorter 6 min workout dedicated to abs twice a week).  There are different 'programs' focused on different areas, for weight loss you do the Lean Program, standard weight loss and muscle gain do the Regular Program, and for those hardcore health-nuts, the Insane Program (which consists of 2 - 1 hour long exercises per day).  Each Program has a different set of workouts per week which you repeat for 3 weeks, followed by a 'Relaxation Week' which is essentially a slightly different order.  After the month of workouts is over, you've finished 1 phase out of 3.  P90X takes 90 days, split into 3 Phases (1 phase per month).  Every phase has a different workout order which is also focused on different areas (Weight Loss, Muscle Gain, etc...)  With the DVD's you also get a small glossy book of about 100 pages detailing the different workouts and the different programs as well as a sample workout to see if you're even ready to start P90X. The second part of P90X, which can also be considered the 'core' (actually the other half of the core) is the nutrition guide that is included.  The Nutrition Guide is a book similar to the one that defines the exercises (about 100 glossy pages) though it details foods you should eat, the amounts, and a number of healthy (and tasty!) recipes.  The guide is split up into 3 phases as well, promoting high protein and low carb/dairy at during Phase 1, and levelling off through to Phase 3 where you have a relatively balanced amount of every food group. So after 1 week where am I?  I've stuck quite close to the nutrition guide (there isn't 'diet food' in here people, it's ACTUALLY food) and done my exercise every day.  I think a lot of the first week is getting into the whole idea and learning the moves performed on the DVD.  Have I lost weight?  No.  Do I feel some definition already starting to poke out?  Absolutely (no pun intended). Tony Horton (the 51-year old hulk that runs the whole thing) is very fun to listen and work along with and the 'diet' really isn't too hard to follow unless all you eat is carbs.  I've tried the gym thing and could not get motivated enough to continue going.  P90X is the first time I've ached from a workout, BEFORE starting my next workout.  For anyone interested, Google 'P90X' or 'BeachBody' to find out more information about this awesome program!

    Read the article

  • Xsigo and Oracle's Storage

    - by Philippe Deverchère
    Xsigo, a virtual network infrastructure provider, has recently been acquired by Oracle. Following this acquisition, one might ask ourselves why it is important to Oracle and how Oracle's storage is going to benefit on the long term from this virtualized infrastructure layer. Well, the first thing to understand is that Virtual Networking addresses both network and storage connectivity. Oracle Virtual Networking, as the Xsigo technology is now called, connects any server to any network and storage, so this is not just about connecting servers to the Internet or Intranet. It is also for a large part connecting servers to NAS and SAN storage. Connecting servers to storage has become increasingly complex in the past few years because of the strong emergence of virtualization at the Operating System level. 50% of enterprise workloads are now virtualized, up from 18% in 2009, resulting in a strong consolidation of various applications in a high density server footprint. At the same time, server I/O capability increased 8x in the last 8 years. All this has pushed IT administrators to multiply the number of I/O connections in the back-end of their physical servers, resulting in a messy and very hard to manage networking infrastructure. Here is a typical view of a rack back-end when no virtual networking is used. We consider that today: - 75% of users have ten or more Ethernet ports per server - 85% of users have two or more SAN ports per server - 58% have had to add connectivity to a server specifically for VMs - 65% consider cable reduction a priority The average is 12 or more ports per server, resulting in an extremely complex infrastructure to manage. What Oracle wants to achieve with its Oracle Virtual Networking offering is pretty simple. The objective is to eliminate the complexity through a dramatic reduction of cabling between servers and storage/networks. It is also to provide a software based management system so that any server can be connected to any network or any storage, on demand, and without physical intervention on the infrastructure. At the end of the day, the picture on the left shows what one wants to get for the back-end of customer's racks: just a couple of connections on each physical server to provide a simple, agile and fast network infrastructure for both storage and networking access. This is exactly what the Oracle Virtual Networking solution does. It transforms a complex, error-prone, difficult to manage and expensive networking infrastructure into a simple, high performance and agile solution for the data center. Practically speaking, and for the sake of simplicity, imagine that each server just hosts a minimal number of physical InfiniBand HCAs (Host Channel Adapter) with two links (for redundancy) onto the Oracle Fabric Interconnect director. Using the Oracle Fabric Manager software, you'll then be able to create virtual NICs and HBAs (called vNIC and vHBA) that will be seen by the servers as standard NICs and HBAs and associate them to networks and storage systems which are physically connected to the back-end of the director through standard Fibre Channel and Ethernet GbE/10GbE ports. In addition to this incredibly simple "at-a-click" connectivity capability, the Oracle Virtual Networking solution offers powerful features such as network isolation, Quality of Service, advanced performance monitoring and non-disruptive reconfiguration, migration and scalability of networking infrastructure. So let's go back now to our initial question: why is Oracle Virtual Networking especially important to Oracle's storage solutions? After all, one could connect any storage in the back-end of the Oracle Fabric Interconnect directors, right? The answer is pretty simple: since Oracle owns both the virtualized networking infrastructure and the storage (ZFS-SA, Pillar Axiom and tape), it is possible to imagine several ways in the future to add value when it comes to connect storage to a virtualized storage network: enhanced storage capabilities, converged management between storage and network, improved diagnostic capabilities and optimized integration resulting in higher performance and unique features/functions. Of course, all this is not going to be done overnight, and future will tell us is which evolutions come first. But there is little doubt that the integration of Xsigo within Oracle is going to create opportunities for Oracle's storage!

    Read the article

  • MaxTotalSizeInBytes - Blind spots in Usage file and Web Analytics Reports

    - by Gino Abraham
    Originally posted on: http://geekswithblogs.net/GinoAbraham/archive/2013/10/28/maxtotalsizeinbytes---blind-spots-in-usage-file-and-web-analytics.aspx http://blogs.msdn.com/b/sharepoint_strategery/archive/2012/04/16/usage-file-and-web-analytics-reports-with-blind-spots.aspx In my previous post (Troubleshooting SharePoint 2010 Web Analytics), I referenced a problem that can occur when exceeding the daily partition size for the LoggingDB, which generates the ULS message “[Partition] has exceeded the max bytes”. Below, I wanted to provide some additional info on this particular issue and help identify some options if this occurs. As an aside, this post only applies if you are missing portions of Usage data - think blind spots on intermittent days or user activity regularly sparse for the afternoon/evening. If this fits your scenario - read on. But if Usage logs are outright missing, go check out my Troubleshooting post first.  Background on the problem:The LoggingDB database has a default maximum size of ~6GB. However, SharePoint evenly splits this total size into fixed sized logical partitions – and the number of partitions is defined by the number of days to retain Usage data (by default 14 days). In this case, 14 partitions would be created to account for the 14 days of retention. If the retention were halved to 7 days, the LoggingDBwould be split into 7 corresponding partitions at twice the size. In other words, the partition size is generally defined as [max size for DB] / [number of retention days].Going back to the default scenario, the “max size” for the LoggingDB is 6200000000 bytes (~6GB) and the retention period is 14 days. Using our formula, this would be [~6GB] / [14 days], which equates to 444858368 bytes (~425MB) per partition per day. Again, if the retention were halved to 7 days (which halves the number of partitions), the resulting partition size becomes [~6GB] / [7 days], or ~850MB per partition.From my experience, when the partition size for any given day is exceeded, the usage logging for the remainder of the day is essentially thrown away because SharePoint won’t allow any more to be written to that day’s partition. The only clue that this is occurring (beyond truncated usage data) is an error such as the following that gets reported in the ULS:04/08/2012 09:30:04.78    OWSTIMER.EXE (0x1E24)    0x2C98    SharePoint Foundation    Health    i0m6     High    Table RequestUsage_Partition12 has 444858368 bytes that has exceeded the max bytes 444858368It’s also worth noting that the exact bytes reported (e.g. ‘444858368’ above) may slightly vary among farms. For example, you may instead see 445226812, 439123456, or something else in the ballpark. The exact number itself doesn't matter, but this error message intends to indicates that the reporting usage has exceeded the partition size for the given day.What it means:The error itself is easy to miss, which can lead to substantial gaps in the reporting data (your mileage may vary) if not identified. At this point, I can only advise to periodically check the ULS logs for this message. Down the road, I plan to explore if [Developing a Custom Health Rule] could be leveraged to identify the issue (If you've ever built Custom Health Rules, I'd be interested to hear about your experiences). Overcoming this issue also poses a challenge, with workaround options including:Lower the retentionBecause the partition size is generally defined as [max size] / [number of retention days], the first option is to lower the number of days to retain the data – the lower the retention, the lower the divisor and thus a bigger partition. For example, halving the retention from 14 to 7 days would halve the number of partitions, but double the partition size to ~850MB (e.g. [6200000000 bytes] / [7 days] = ~850GB partitions). Lowering it to 2 days would result in two ~3GB partitions… and so on.Recreate the LoggingDB with an increased sizeThe property MaxTotalSizeInBytes is exposed by OM code for the SPUsageDefinition object and can be updated with the example PowerShell snippet below. However, updating this value has no immediate impact because this size only applies when creating a LoggingDB. Therefore, you must create a newLoggingDB for the Usage Service Application. The gotcha: this effectively deletes all prior Usage databecause the Usage Service Application can only have a single LoggingDB.Here is an example snippet to update the "Page Requests" Usage Definition:$def=Get-SPUsageDefinition -Identity "page requests" $def.MaxTotalSizeInBytes=12400000000 $def.update()Create a new Logging database and attach to the Usage Service Application using the following command: Get-spusageapplication | Set-SPUsageApplication -DatabaseServer <dbServer> -DatabaseName <newDBname> Updated (5/10/2012): Once the new database has been created, you can confirm the setting has truly taken by running the following SQL Query (be sure to replace the database name in the following query with the name provided in the PowerShell above)SELECT * FROM [WSS_UsageApplication].[dbo].[Configuration] WITH (nolock) WHERE ConfigName LIKE 'Max Total Bytes - RequestUsage'

    Read the article

  • WebCenter Customer Spotlight: Hyundai Motor Company

    - by me
    Author: Peter Reiser - Social Business Evangelist, Oracle WebCenter  Solution SummaryHyundai Motor Company is one of the world’s fastest-growing car manufacturers, ranked as the fifth-largest in 2011. The company also operates the world’s largest integrated automobile manufacturing facility in Ulsan, Republic of Korea, which can produce 1.6 million units per year. They  undertook a project to improve business efficiency and reinforce data security by centralizing the company’s sales, financial, and car manufacturing documents into a single repository. Hyundai Motor Company chose Oracle Exalogic, Oracle Exadata, Oracle WebLogic Sever, and Oracle WebCenter Content 11g, as they provided better performance, stability, storage, and scalability than their competitors.  Hyundai Motor Company cut the overall time spent each day on document-related work by around 85%, saved more than US$1 million in paper and printing costs, laid the foundation for a smart work environment, and supported their future growth in the competitive car industry. Company OverviewHyundai Motor Company is one of the world’s fastest-growing car manufacturers, ranked as the fifth-largest in 2011. The company also operates the world’s largest integrated automobile manufacturing facility in Ulsan, Republic of Korea, which can produce 1.6 million units per year. The company strives to enhance its brand image and market recognition by continuously improving the quality and design of its cars. Business Challenges To maximize the company’s growth potential, Hyundai Motor Company undertook a project to improve business efficiency and reinforce data security by centralizing the company’s sales, financial, and car manufacturing documents into a single repository. Specifically, they wanted to: Introduce a smart work environment to improve staff productivity and efficiency, and take advantage of rapid company growth due to new, enhanced car designs Replace a legacy document system managed by individual staff to improve collaboration, the visibility of corporate documents, and sharing of work-related files between employees Improve the security and storage of documents containing corporate intellectual property, and prevent intellectual property loss when staff leaves the company Eliminate delays when downloading files from the central server to a PC Build a large, single document repository to more efficiently manage and share data between 30,000 staff at the company’s headquarters Establish a scalable system that can be extended to Hyundai offices around the world Solution DeployedAfter conducting a large-scale benchmark test, Hyundai Motor Company chose Oracle Exalogic, Oracle Exadata, Oracle WebLogic Sever, and Oracle WebCenter Content 11g, as they provided better performance, stability, storage, and scalability than their competitors. Business Results Lowered the overall time spent each day on all document-related work by approximately 85%—from 4.5 hours to around 42 minutes on an average day Saved more than US$1 million per year in printer, paper, and toner costs, and laid the foundation for a completely paperless environment Reduced staff’s time spent requesting and receiving documents about car sales or designs from supervisors by 50%, by storing and managing all documents across the corporation in a single repository Cut the time required to draft new-car manufacturing, sales, and design documents by 20%, by allowing employees to reference high-quality data, such as marketing strategy and product planning documents already in the system Enhanced staff productivity at company headquarters by 9% by reducing the document-related tasks of 30,000 administrative and research and development staff Ensured the system could scale to hold 3 petabytes of car sales, manufacturing, and design data by 2013 and be deployed at branches worldwide We chose Oracle Exalogic, Oracle Exadata, and Oracle WebCenter Content to support our new document-centralization system over their competitors as Oracle offers stable storage for petabytes of data and high processing speeds. We have cut the overall time spent each day on document-related work by around 85%, saved more than US$1 million in paper and printing costs, laid the foundation for a smart work environment, and supported our future growth in the competitive car industry. Kang Tae-jin, Manager, General Affairs Team, Hyundai Motor Company Additional Information Hyundai Motor Company Customer Snapshot Oracle WebCenter Content

    Read the article

  • Upcoming Carbon Tax in South Africa

    - by Evelyn Neumayr
    By Elena Avesani, Principal Product Strategy Manager, Oracle In 2012, the South Africa National Treasury announced the plan to impose a carbon tax to cut carbon emissions that are blamed for climate change. South Africa is ranked among the top 20 countries measured by absolute carbon dioxide emissions, with emissions per capita in the region of 10 metric tons per annum and over 90% of South Africa's energy produced by burning fossil fuels. The top 40 largest companies in the country are responsible for 207 million tons of carbon dioxide, directly emitting 20 percent of South Africa’s carbon output. The legislation, originally scheduled to be implemented from January 2015 to 31 December 2019, is now delayed to January 2016. It will levy a carbon tax of R120 (US$11) per ton of CO2, rising then by 10 percent a year until 2020, while all sectors bar electricity will be able to claim additional relief of at least 10 percent. The South African treasury proposed a 60 percent tax-free threshold on emissions for all sectors, including electricity, petroleum, iron, steel and aluminum. Oracle Environmental Accounting and Reporting (EA&R) supports these needs and guarantees consistency across organizations in how data is collected, retained, controlled, consolidated and used in calculating and reporting emissions inventory. EA&R also enables companies to develop an enterprise-wide data view that includes all 5 of the key sustainability categories: carbon emissions, energy, water, materials and waste. Thanks to its native integration with Oracle E-Business Suite and JD Edwards EnterpriseOne ERP Financials and Inventory Systems and the capability of capturing environmental data across business silos, Oracle Environmental Accounting and Reporting is uniquely positioned to support a strategic approach to carbon management that drives business value. Sources: Normal 0 false false false EN-US X-NONE X-NONE /* Style Definitions */ table.MsoNormalTable {mso-style-name:"Table Normal"; mso-tstyle-rowband-size:0; mso-tstyle-colband-size:0; mso-style-noshow:yes; mso-style-priority:99; mso-style-qformat:yes; mso-style-parent:""; mso-padding-alt:0in 5.4pt 0in 5.4pt; mso-para-margin:0in; mso-para-margin-bottom:.0001pt; mso-pagination:widow-orphan; font-size:11.0pt; font-family:"Calibri","sans-serif"; mso-ascii-font-family:Calibri; mso-ascii-theme-font:minor-latin; mso-fareast-font-family:"Times New Roman"; mso-fareast-theme-font:minor-fareast; mso-hansi-font-family:Calibri; mso-hansi-theme-font:minor-latin; mso-bidi-font-family:"Times New Roman"; mso-bidi-theme-font:minor-bidi;} African Utility Week BDlive Normal 0 false false false EN-US X-NONE X-NONE /* Style Definitions */ table.MsoNormalTable {mso-style-name:"Table Normal"; mso-tstyle-rowband-size:0; mso-tstyle-colband-size:0; mso-style-noshow:yes; mso-style-priority:99; mso-style-qformat:yes; mso-style-parent:""; mso-padding-alt:0in 5.4pt 0in 5.4pt; mso-para-margin:0in; mso-para-margin-bottom:.0001pt; mso-pagination:widow-orphan; font-size:11.0pt; font-family:"Calibri","sans-serif"; mso-ascii-font-family:Calibri; mso-ascii-theme-font:minor-latin; mso-fareast-font-family:"Times New Roman"; mso-fareast-theme-font:minor-fareast; mso-hansi-font-family:Calibri; mso-hansi-theme-font:minor-latin; mso-bidi-font-family:"Times New Roman"; mso-bidi-theme-font:minor-bidi;}

    Read the article

  • Faster Memory Allocation Using vmtasks

    - by Steve Sistare
    You may have noticed a new system process called "vmtasks" on Solaris 11 systems: % pgrep vmtasks 8 % prstat -p 8 PID USERNAME SIZE RSS STATE PRI NICE TIME CPU PROCESS/NLWP 8 root 0K 0K sleep 99 -20 9:10:59 0.0% vmtasks/32 What is vmtasks, and why should you care? In a nutshell, vmtasks accelerates creation, locking, and destruction of pages in shared memory segments. This is particularly helpful for locked memory, as creating a page of physical memory is much more expensive than creating a page of virtual memory. For example, an ISM segment (shmflag & SHM_SHARE_MMU) is locked in memory on the first shmat() call, and a DISM segment (shmflg & SHM_PAGEABLE) is locked using mlock() or memcntl(). Segment operations such as creation and locking are typically single threaded, performed by the thread making the system call. In many applications, the size of a shared memory segment is a large fraction of total physical memory, and the single-threaded initialization is a scalability bottleneck which increases application startup time. To break the bottleneck, we apply parallel processing, harnessing the power of the additional CPUs that are always present on modern platforms. For sufficiently large segments, as many of 16 threads of vmtasks are employed to assist an application thread during creation, locking, and destruction operations. The segment is implicitly divided at page boundaries, and each thread is given a chunk of pages to process. The per-page processing time can vary, so for dynamic load balancing, the number of chunks is greater than the number of threads, and threads grab chunks dynamically as they finish their work. Because the threads modify a single application address space in compressed time interval, contention on locks protecting VM data structures locks was a problem, and we had to re-scale a number of VM locks to get good parallel efficiency. The vmtasks process has 1 thread per CPU and may accelerate multiple segment operations simultaneously, but each operation gets at most 16 helper threads to avoid monopolizing CPU resources. We may reconsider this limit in the future. Acceleration using vmtasks is enabled out of the box, with no tuning required, and works for all Solaris platform architectures (SPARC sun4u, SPARC sun4v, x86). The following tables show the time to create + lock + destroy a large segment, normalized as milliseconds per gigabyte, before and after the introduction of vmtasks: ISM system ncpu before after speedup ------ ---- ------ ----- ------- x4600 32 1386 245 6X X7560 64 1016 153 7X M9000 512 1196 206 6X T5240 128 2506 234 11X T4-2 128 1197 107 11x DISM system ncpu before after speedup ------ ---- ------ ----- ------- x4600 32 1582 265 6X X7560 64 1116 158 7X M9000 512 1165 152 8X T5240 128 2796 198 14X (I am missing the data for T4 DISM, for no good reason; it works fine). The following table separates the creation and destruction times: ISM, T4-2 before after ------ ----- create 702 64 destroy 495 43 To put this in perspective, consider creating a 512 GB ISM segment on T4-2. Creating the segment would take 6 minutes with the old code, and only 33 seconds with the new. If this is your Oracle SGA, you save over 5 minutes when starting the database, and you also save when shutting it down prior to a restart. Those minutes go directly to your bottom line for service availability.

    Read the article

  • Using WKA in Large Coherence Clusters (Disabling Multicast)

    - by jpurdy
    Disabling hardware multicast (by configuring well-known addresses aka WKA) will place significant stress on the network. For messages that must be sent to multiple servers, rather than having a server send a single packet to the switch and having the switch broadcast that packet to the rest of the cluster, the server must send a packet to each of the other servers. While hardware varies significantly, consider that a server with a single gigabit connection can send at most ~70,000 packets per second. To continue with some concrete numbers, in a cluster with 500 members, that means that each server can send at most 140 cluster-wide messages per second. And if there are 10 cluster members on each physical machine, that number shrinks to 14 cluster-wide messages per second (or with only mild hyperbole, roughly zero). It is also important to keep in mind that network I/O is not only expensive in terms of the network itself, but also the consumption of CPU required to send (or receive) a message (due to things like copying the packet bytes, processing a interrupt, etc). Fortunately, Coherence is designed to rely primarily on point-to-point messages, but there are some features that are inherently one-to-many: Announcing the arrival or departure of a member Updating partition assignment maps across the cluster Creating or destroying a NamedCache Invalidating a cache entry from a large number of client-side near caches Distributing a filter-based request across the full set of cache servers (e.g. queries, aggregators and entry processors) Invoking clear() on a NamedCache The first few of these are operations that are primarily routed through a single senior member, and also occur infrequently, so they usually are not a primary consideration. There are cases, however, where the load from introducing new members can be substantial (to the point of destabilizing the cluster). Consider the case where cluster in the first paragraph grows from 500 members to 1000 members (holding the number of physical machines constant). During this period, there will be 500 new member introductions, each of which may consist of several cluster-wide operations (for the cluster membership itself as well as the partitioned cache services, replicated cache services, invocation services, management services, etc). Note that all of these introductions will route through that one senior member, which is sharing its network bandwidth with several other members (which will be communicating to a lesser degree with other members throughout this process). While each service may have a distinct senior member, there's a good chance during initial startup that a single member will be the senior for all services (if those services start on the senior before the second member joins the cluster). It's obvious that this could cause CPU and/or network starvation. In the current release of Coherence (3.7.1.3 as of this writing), the pure unicast code path also has less sophisticated flow-control for cluster-wide messages (compared to the multicast-enabled code path), which may also result in significant heap consumption on the senior member's JVM (from the message backlog). This is almost never a problem in practice, but with sufficient CPU or network starvation, it could become critical. For the non-operational concerns (near caches, queries, etc), the application itself will determine how much load is placed on the cluster. Applications intended for deployment in a pure unicast environment should be careful to avoid excessive dependence on these features. Even in an environment with multicast support, these operations may scale poorly since even with a constant request rate, the underlying workload will increase at roughly the same rate as the underlying resources are added. Unless there is an infrastructural requirement to the contrary, multicast should be enabled. If it can't be enabled, care should be taken to ensure the added overhead doesn't lead to performance or stability issues. This is particularly crucial in large clusters.

    Read the article

  • How to improve WinForms MSChart performance?

    - by Marcel
    Hi all, I have created some simple charts (of type FastLine) with MSChart and update them with live data, like below: . To do so, I bind an observable collection of a custom type to the chart like so: // set chart data source this._Chart.DataSource = value; //is of type ObservableCollection<SpectrumLevels> //define x and y value members for each series this._Chart.Series[0].XValueMember = "Index"; this._Chart.Series[1].XValueMember = "Index"; this._Chart.Series[0].YValueMembers = "Channel0Level"; this._Chart.Series[1].YValueMembers = "Channel1Level"; // bind data to chart this._Chart.DataBind(); //lasts 1.5 seconds for 8000 points per series At each refresh, the dataset completely changes, it is not a scrolling update! With a profiler I have found that the DataBind() call takes about 1.5 seconds. The other calls are negligible. How can I make this faster? Should I use another type than ObservableCollection? An array probably? Should I use another form of data binding? Is there some tweak for the MSChart that I may have missed? Should I use a sparsed set of date, having one value per pixel only? Have I simply reached the performance limit of MSCharts? From the type of the application to keep it "fluent", we should have multiple refreshes per second. Thanks for any hints!

    Read the article

  • We have multiple app servers running against a single database. How do I ensure that each row in a q

    - by Dave
    We have about 7 app servers running .NET windows services that ping a single sql server 2005 queue table and fetch a fixed amount of records to process at fixed intervals. The amount of records to process and the amount of time between fetches are both configurable and are initially set to 100 and 30 seconds initially. Currently, my queue table has an int status column which can be either "Ready, Processing, Complete, Error". The proc that fetches the records has a sql transaction with the following code inside the transaction: 1) Fetch x number of records into temp table where the status is "Ready". The select uses a holdlock hint 2) Update the status on those records in the Queue table to "Processing" The .NET services do some processing that may take seconds or even minutes per record. Another proc is called per record that simply updates the status to "Complete". The update proc has no transaction as I'm leaning on the implicit transaction as part of the update clause here. I don't know the traffic exceptions for this but figure it will be under 10k records per day. Is this the best way to handle this scenario? If so, are there any details that I've left out, such as a hint here or there? Thanks! Dave

    Read the article

  • Which key:value store to use with Python?

    - by Kurt
    So I'm looking at various key:value (where value is either strictly a single value or possibly an object) stores for use with Python, and have found a few promising ones. I have no specific requirement as of yet because I am in the evaluation phase. I'm looking for what's good, what's bad, what are the corner cases these things handle well or don't, etc. I'm sure some of you have already tried them out so I'd love to hear your findings/problems/etc. on the various key:value stores with Python. I'm looking primarily at: memcached - http://www.danga.com/memcached/ python clients: http://pypi.python.org/pypi/python-memcached/1.40 http://www.tummy.com/Community/software/python-memcached/ CouchDB - http://couchdb.apache.org/ python clients: http://code.google.com/p/couchdb-python/ Tokyo Tyrant - http://1978th.net/tokyotyrant/ python clients: http://code.google.com/p/pytyrant/ Lightcloud - http://opensource.plurk.com/LightCloud/ Based on Tokyo Tyrant, written in Python Redis - http://code.google.com/p/redis/ python clients: http://pypi.python.org/pypi/txredis/0.1.1 MemcacheDB - http://memcachedb.org/ So I started benchmarking (simply inserting keys and reading them) using a simple count to generate numeric keys and a value of "A short string of text": memcached: CentOS 5.3/python-2.4.3-24.el5_3.6, libevent 1.4.12-stable, memcached 1.4.2 with default settings, 1 gig memory, 14,000 inserts per second, 16,000 seconds to read. No real optimization, nice. memcachedb claims on the order of 17,000 to 23,000 inserts per second, 44,000 to 64,000 reads per second. I'm also wondering how the others stack up speed wise.

    Read the article

  • Memory mapped files and "soft" page faults. Unavoidable?

    - by Robert Oschler
    I have two applications (processes) running under Windows XP that share data via a memory mapped file. Despite all my efforts to eliminate per iteration memory allocations, I still get about 10 soft page faults per data transfer. I've tried every flag there is in CreateFileMapping() and CreateFileView() and it still happens. I'm beginning to wonder if it's just the way memory mapped files work. If anyone there knows the O/S implementation details behind memory mapped files I would appreciate comments on the following theory: If two processes share a memory mapped file and one process writes to it while another reads it, then the O/S marks the pages written to as invalid. When the other process goes to read the memory areas that now belong to invalidated pages, this causes a soft page fault (by design) and the O/S knows to reload the invalidated page. Also, the number of soft page faults is therefore directly proportional to the size of the data write. My experiments seem to bear out the above theory. When I share data I write one contiguous block of data. In other words, the entire shared memory area is overwritten each time. If I make the block bigger the number of soft page faults goes up correspondingly. So, if my theory is true, there is nothing I can do to eliminate the soft page faults short of not using memory mapped files because that is how they work (using soft page faults to maintain page consistency). What is ironic is that I chose to use a memory mapped file instead of a TCP socket connection because I thought it would be more efficient. Note, if the soft page faults are harmless please note that. I've heard that at some point if the number is excessive, the system's performance can be marred. If soft page faults intrinsically are not significantly harmful then if anyone has any guidelines as to what number per second is "excessive" I'd like to hear that. Thanks.

    Read the article

  • Figuring out the performance limitation of an ADC on a PIC microcontroller

    - by AKE
    I'm spec-ing the suitability of a microcontroller like PIC for an analog-to-digital application. This would be preferable to using external A/D chips. To do that, I've had to run through some computations, pulling the relevant parameters from the datasheets. I'm not sure I've got it right -- would appreciate a check! Here's the simplest example: PIC10F220 is the simplest possible PIC with an ADC. Runs at clock speed of 8MHz. Has an instruction cycle of 0.5us (4 clock steps per instruction) So: Taking Tacq = 6.06 us (acquisition time for ADC, assume chip temp. = 50*C) [datasheet p34] Taking Fosc = 8MHz (? clock speed) Taking divisor = 4 (4 clock steps per CPU instruction) This gives TAD = 0.5us (TAD = 1/(Fosc/divisor) ) Conversion time is 13*TAD [datasheet p31] This gives conversion time 6.5us ADC duration is then 12.56 us [? Tacq + 13*TAD] Assuming at least 2 instructions for load/store: This is another 1 us [0.5 us per instruction] Which would give max sampling rate of 73.7 ksps (1/13.56) Supposing 8 more instructions for real-time processing: This is another 4 us Thus, total ADC/handling time = 17.56us (12.56us + 1us + 4us) So expected upper sampling rate is 56.9 ksps. Nyquist frequency for this sampling rate is therefore 28 kHz. If this is right, it suggests the (theoretical) performance suitability of this chip's A/D is for signals that are bandlimited to 28 kHz. Is this a correct interpretation of the information given in the data sheet? Any pointers would be much appreciated! AKE

    Read the article

  • Several Small, Specific, MySQL Query Cache Questions

    - by Robbie
    I've look all over the web and in the questions asked here about MySQL caching and most of them seem very non-specific about a couple of questions that I have about performance and MySQL query caching. Specifically I want answers to these questions, assume for all questions that I have the query cache enabled and it is of type 2, or "DEMAND": Is the query cache per table, per database, or per server? Meaning if I have the cache size set to X and have T tables and D databases will I be caching TX, DX, or X amount of data? If I have table T1 which I regularly use the SQL_CACHE hint on for SELECT queries and table T2 which I never do, when I query T2 with a SELECT query will it check through the cache first before performing the query? *Note: I don't want to use the SQL_NO_CACHE for all T2 queries.* Assume the same situation as in question 2. If I alter (INSERT, DELETE) table T2 will any processing be done on the cache? For answers to 2 and 3, is this processing time negligible if T2 is constantly being altered and is the target of a majority of my SELECT queries?

    Read the article

  • Figuring out the Nyquist performance limitation of an ADC on an example PIC microcontroller

    - by AKE
    I'm spec-ing the suitability of a dsPIC microcontroller for an analog-to-digital application. This would be preferable to using dedicated A/D chips and a separate dedicated DSP chip. To do that, I've had to run through some computations, pulling the relevant parameters from the datasheets. I'm not sure I've got it right -- would appreciate a check! (EDITED NOTE: The PIC10F220 in the example below was selected ONLY to walk through a simple example to check that I'm interpreting Tacq, Fosc, TAD, and divisor correctly in working through this sort of Nyquist analysis. The actual chips I'm considering for the design are the dsPIC33FJ128MC804 (with 16b A/D) or dsPIC30F3014 (with 12b A/D).) A simple example: PIC10F220 is the simplest possible PIC with an ADC Runs at clock speed of 8MHz. Has an instruction cycle of 0.5us (4 clock steps per instruction) So: Taking Tacq = 6.06 us (acquisition time for ADC, assume chip temp. = 50*C) [datasheet p34] Taking Fosc = 8MHz (? clock speed) Taking divisor = 4 (4 clock steps per CPU instruction) This gives TAD = 0.5us (TAD = 1/(Fosc/divisor) ) Conversion time is 13*TAD [datasheet p31] This gives conversion time 6.5us ADC duration is then 12.56 us [? Tacq + 13*TAD] Assuming at least 2 instructions for load/store: This is another 1 us [0.5 us per instruction] Which would give max sampling rate of 73.7 ksps (1/13.56) Supposing 8 more instructions for real-time processing: This is another 4 us Thus, total ADC/handling time = 17.56us (12.56us + 1us + 4us) So expected upper sampling rate is 56.9 ksps. Nyquist frequency for this sampling rate is therefore 28 kHz. If this is right, it suggests the (theoretical) performance suitability of this chip's A/D is for signals that are bandlimited to 28 kHz. Is this a correct interpretation of the information given in the data sheet in obtaining the Nyquist performance limit? Any opinions on the noise susceptibility of ADCs in PIC / dsPIC chips would be much appreciated! AKE

    Read the article

  • Internet Explorer cannot 'fully' load ActiveX Control

    - by K Browne
    Context I am migrating an installer for an ActiveX control from Per-Machine to Per-User. I did this by programming the installer write to HKCU\Software\Classes instead of HKLM\Software\Classes. Problem On my machine (Windows 7 with UAC Enabled), the ActiveX control successfully loads. On the other windows 7 test machines (one with UAC enabled, one with UAC disabled), the control 'partially' loads. What is Partially? When a user visits a page with the ActiveX control, Internet Explorer displays a warning message in a yellow bar on the top of the window. If you click the 'Run add-on' button in the bar, the control becomes visible and begins to run, but Javascript code that tries to access properties of the control return the error: Library not registered. Differences between machines On the dev machine reads from HKCR\CLSID\<GUID> succeed while on the test machines these reads fail. Reads from HKCU succeed on both dev and test machines. Reads from HKLM fail on both test and dev machines. (I collected reads using Sysinternals Process Monitor) Strangely, the keys that Internet Explorer fails to read are clearly visible if I use regedit to view HKCR\CLSID\<GUID> on the test machines. Question What can I do to get the per-user control to load on the test machines? What could cause this difference between the dev machine and the test machines? Why can I see the key in HKCR with RegEdit but Internet Explorer cannot see the key? Any help is appreciated. Thank you.

    Read the article

  • MySQL - How do I inner join sorting the joined data

    - by Gary
    I'm trying to write a report which will join a person, their work, and their hourly wage at the time of work. I cannot seem to figure out the best way to join the person's cost when the date is less than the date of the work. Let's say a person cost $30 per hour at the start of the year then got a $10 raise o Feb 5 and another on Mar 1. 01/01/2010 $30.00 (per hour) 02/05/2010 $40.00 03/01/2010 $45.00 The person put in hours several days which span the rasies. 01/05/2010 10 hours (should be at $30/hr) 01/27/2010 5 hours (again at $30) 02/10/2010 10 hours (at $40/hr) 03/03/2010 5 hours (at $45/hr) I'm trying to write one SQL statement which will pull the hours, the cost per hour, and the hours*cost. The cost is the hourly rate last entered into the system so the cost date is less than the work date, ordered by cost date limit 1. SELECT person.id, person.name, work.hours, person_costs.value, work.hours * person_costs.value AS value FROM person INNER JOIN work ON (person.id = work.person_id) INNER JOIN person_costs ON (person.id = person_costs.person_id AND person_costs.date < work.date) WHERE person.id = 1234 ORDER BY work.date ASC The problem I'm having, the person_costs isn't ordered by date in descending order. It's pulling out "any" value (naturally sorted by record position) which matches the condition. How do I select the first person_cost value which is older than the work date? Thanks!

    Read the article

  • MySQL MyISAM table performance... painfully, painfully slow

    - by Salman A
    I've got a table structure that can be summarized as follows: pagegroup * pagegroupid * name has 3600 rows page * pageid * pagegroupid * data references pagegroup; has 10000 rows; can have anything between 1-700 rows per pagegroup; the data column is of type mediumtext and the column contains 100k - 200kbytes data per row userdata * userdataid * pageid * column1 * column2 * column9 references page; has about 300,000 rows; can have about 1-50 rows per page The above structure is pretty straight forwad, the problem is that that a join from userdata to page group is terribly, terribly slow even though I have indexed all columns that should be indexed. The time needed to run a query for such a join (userdata inner_join page inner_join pagegroup) exceeds 3 minutes. This is terribly slow considering the fact that I am not selecting the data column at all. Example of the query that takes too long: SELECT userdata.column1, pagegroup.name FROM userdata INNER JOIN page USING( pageid ) INNER JOIN pagegroup USING( pagegroupid ) Please help by explaining why does it take so long and what can i do to make it faster. Edit #1 Explain returns following gibberish: id select_type table type possible_keys key key_len ref rows Extra 1 SIMPLE userdata ALL pageid 372420 1 SIMPLE page eq_ref PRIMARY,pagegroupid PRIMARY 4 topsecret.userdata.pageid 1 1 SIMPLE pagegroup eq_ref PRIMARY PRIMARY 4 topsecret.page.pagegroupid 1 Edit #2 SELECT u.field2, p.pageid FROM userdata u INNER JOIN page p ON u.pageid = p.pageid; /* 0.07 sec execution, 6.05 sec fecth */ id select_type table type possible_keys key key_len ref rows Extra 1 SIMPLE u ALL pageid 372420 1 SIMPLE p eq_ref PRIMARY PRIMARY 4 topsecret.u.pageid 1 Using index SELECT p.pageid, g.pagegroupid FROM page p INNER JOIN pagegroup g ON p.pagegroupid = g.pagegroupid; /* 9.37 sec execution, 60.0 sec fetch */ id select_type table type possible_keys key key_len ref rows Extra 1 SIMPLE g index PRIMARY PRIMARY 4 3646 Using index 1 SIMPLE p ref pagegroupid pagegroupid 5 topsecret.g.pagegroupid 3 Using where Moral of the story Keep medium/long text columns in a separate table if you run into performance problems such as this one.

    Read the article

  • How to improve performance of non-scalar aggregations on denormalized tables

    - by The Lazy DBA
    Suppose we have a denormalized table with about 80 columns, and grows at the rate of ~10 million rows (about 5GB) per month. We currently have 3 1/2 years of data (~400M rows, ~200GB). We create a clustered index to best suit retrieving data from the table on the following columns that serve as our primary key... [FileDate] ASC, [Region] ASC, [KeyValue1] ASC, [KeyValue2] ASC ... because when we query the table, we always have the entire primary key. So these queries always result in clustered index seeks and are therefore very fast, and fragmentation is kept to a minimum. However, we do have a situation where we want to get the most recent FileDate for every Region, typically for reports, i.e. SELECT [Region] , MAX([FileDate]) AS [FileDate] FROM HugeTable GROUP BY [Region] The "best" solution I can come up to this is to create a non-clustered index on Region. Although it means an additional insert on the table during loads, the hit isn't minimal (we load 4 times per day, so fewer than 100,000 additional index inserts per load). Since the table is also partitioned by FileDate, results to our query come back quickly enough (200ms or so), and that result set is cached until the next load. However I'm guessing that someone with more data warehousing experience might have a solution that's more optimal, as this, for some reason, doesn't "feel right".

    Read the article

  • Log4Net GetLogger creates rolling files even for the unreferenced files

    - by ybastiand
    Hi, I have a C# solution that contains three executables. I have each of these three executables sharing the same log4net configuration file. At startup of each of the executable, they retrieve a logger (one logger per executable, as per configuration file further below). When one of the executable performs Log.GetLogger(), it creates all the rolling files instead of only the one rolling file that is referred to as appender-ref in the executable's logger configuration. For instance, when I startup my sending daemon executable, it performs Log.GetLogger("SendingDaemonLogger") which creates 3 files Log/RuleScheduler.txt, Log/NotificationGenerator.txt and Log/NotificationSender.txt instead of only the desired Log/NotificationSender.txt. Then when I startup another of the executables, for instance the rule scheduler daemon, this other process cannot write in Log/RuleScheduler.txt because it has been created and locked by the sending daemon process. I am guessing that there may be three different solutions to my problem: The GetLogger should only create the rolling file appenders that are referenced in the config I should have one config file per executable, this way each config file could list only one rolling file appender and starting each of the executable would not create the rolling files of the other daemons. I am however reluctant to do this because some of the configuration (SMTP appender, console appender) is shared between the daemons and I don't want to have duplicate copies to maintain. Unless there is a way to have a config file including another one? Maybe there is a way to configure the rolling file so that concurrent access across processes is allowed? This solution still isn't perfect in my opinion because any of the daemons should not be creating the rolling files of some other daemons. Thanks in advance for your help! I have difficulties for posting the config file properly here (this website interprets as HTML). Please go to the following link for seeing my log4net configuration file: log4Net configuration file

    Read the article

  • Serial: write() throttling?

    - by damian
    Hi everyone, I'm working on a project sending serial data to control animation of LED lights, which need to stay in sync with a sound engine. There seems to be a large serial write buffer (OSX (POSIX) + FTDI chipset usb serial device), so without manually restricting the transmission rate, the animation system can get several seconds ahead of the serial transmission. Currently I'm manually restricting the serial write speed to the baudrate (8N1 = 10 bytes serial frame per 8 bytes data, 19200 bps serial - 1920 bytes per second max), but I am having a problem with the sound drifting out of sync over time - it starts fine, but after 10 minutes there's a noticeable (100ms+) lag between the sound and the lights. This is the code that's restricting the serial write speed (called once per animation frame, 'elapsed' is the duration of the current frame, 'baudrate' is the bps (19200)): void BufferedSerial::update( float elapsed ) { baud_timer += elapsed; if ( bytes_written > 1024 ) { // maintain baudrate float time_should_have_taken = (float(bytes_written)*10)/float(baudrate); float time_actually_took = baud_timer; // sleep if we have > 20ms lag between serial transmit and our write calls if ( time_should_have_taken-time_actually_took > 0.02f ) { float sleep_time = time_should_have_taken - time_actually_took; int sleep_time_us = sleep_time*1000.0f*1000.0f; //printf("BufferedSerial::update sleeping %i ms\n", sleep_time_us/1000 ); delayUs( sleep_time_us ); // subtract 128 bytes bytes_written -= 128; // subtract the time it should have taken to write 128 bytes baud_timer -= (float(128)*10)/float(baudrate); } } } Clearly there's something wrong, somewhere. A much better approach would be to be able to determine the number of bytes currently in the transmit queue, and try and keep that below a fixed threshold. Any advice appreciated.

    Read the article

  • How can I speed up Subversion checkins? (Using ANKH, latest, Visual Studio 2010)

    - by Timothy Khouri
    I've started working on a new web project with some friends... we are using the latest Subversion server (installed last week), the latest version of ANKH. My web project is a whapping 1.5 megabytes (that's with all images, css files, dll's after compiling, pdb files... etc). Checking in even super small changes (literally adding the letter "x" to a few files for testing)... takes FOREVER! (about 10 seconds - I almost killed myself). The ANKH client is measuring in BYTES PER SECOND ... BYTES? per second... I must be doing something wrong. Does anyone what config file has a joke totallyMessWithPeople=true so that I can turn that off or something? Oh, also, changing one "big" file of a super 10k gains speed up to nearly the speed of light (which is apparently 857 bytes per second). Help me obi wan kenobi, your my only hope! EDIT: As a note... my real work project that uses Visual Source Safe 2005 (I know, ouch) uploads files at about 200-500kbps from this very same computer/internet connection.

    Read the article

< Previous Page | 47 48 49 50 51 52 53 54 55 56 57 58  | Next Page >