Search Results

Search found 25 results on 1 pages for 'supercomputer'.

facebook twitter digg google delicious technorati stumbleupon myspace wordpress linkedin gmail igoogle windows live tumblr viadeo yahoo buzz yahoo mail yahoo bookmarks favorites email print

Page 1/1 | 1 

  • Things to do with supercomputer.

    - by BigBoss
    HI, Guys I wanted to create SuperComputer out of my 3 old pcs soon. There are bunch of techniquies available on net for doing this and I will be using one of them. What I would like to get suggestion is what is next !! Many of them suggest some mathematical calculation or some graphic image rendering, which still is to me so boring and not inspiring at all. Please try to suggest something out of box. I would post results of it when I am done. Cheers !

    Read the article

  • 1tera flop cluster?

    - by Adobe
    I want to buy a $40000 1 tera flop cluster to keep it in a room. What are the standard configurations? Cluster is supposed to do molecular dynamics simulations on biological systems. I'm proposed a 4 pc with 8 cores each by the selling company I'm deadling with. It looks like I also need infiniband. Does some one has an experience -- what phisical memory should I buy etc? I know things change very quickly... Still there might be a point or two to state. Edit: OS is supposed to be linux, application is gromacs.

    Read the article

  • A Real-Time HPC Approach for Optimizing Multicore Architectures

    Complex math is at the heart of many of the biggest technical challenges. With multicore processors, the type of calculations that would have required a supercomputer can now be performed in real-time, embedded environments. High-performance computing - Supercomputer - Real-time computing - Operating system - Companies

    Read the article

  • Super Computer Built from Raspberry Pi Boards and LEGO Bricks

    - by Jason Fitzpatrick
    It was only a matter of time before someone chained together dozens of Raspberry Pi boards into a serviceable super computer; read on to see how a team of Southampton scientists built a 64-core machine using them. Image courtesy of Simon Cox and the University of Southampton. From the University of South Hampton press release: Professor Cox comments: “As soon as we were able to source sufficient Raspberry Pi computers we wanted to see if it was possible to link them together into a supercomputer. We installed and built all of the necessary software on the Pi starting from a standard Debian Wheezy system image and we have published a guide so you can build your own supercomputer.” The racking was built using Lego with a design developed by Simon and James, who has also been testing the Raspberry Pi by programming it using free computer programming software Python and Scratch over the summer. The machine, named “Iridis-Pi” after the University’s Iridis supercomputer, runs off a single 13 Amp mains socket and uses MPI (Message Passing Interface) to communicate between nodes using Ethernet. The whole system cost under £2,500 (excluding switches) and has a total of 64 processors and 1Tb of memory (16Gb SD cards for each Raspberry Pi). Professor Cox uses the free plug-in ‘Python Tools for Visual Studio’ to develop code for the Raspberry Pi. How to Get Pro Features in Windows Home Versions with Third Party Tools HTG Explains: Is ReadyBoost Worth Using? HTG Explains: What The Windows Event Viewer Is and How You Can Use It

    Read the article

  • Téléchargez gratuitement l'ebook sur le développement d'applications 'Threaded' qui utilisent le har

    Téléchargez gratuitement l'ebook sur le développement d'applications ?Threaded' Les logiciels de développement Intel® Parallel Studio accélèrent le développement d'applications ?Threaded' qui utilisent le hardware des utilisateurs finaux, depuis le ?'supercomputer'' jusqu'à l'ordinateur portable ou les mobiles. Optimisez la performance de votre application sur architecture Intel® et obtenez plus des derniers processeurs multi-coeurs d'Intel®. Depuis la manière dont les produits fonctionnent ensemble jusqu'à leurs jeux de fonctionnalités uniques, le Threading est maintenant plus facile et plus viable que jamais. Les outils sont optimisés donc les novices peuvent facilement se former et les développeurs expérimentés peuvent aisément ...

    Read the article

  • Téléchargez gratuitement l'ebook sur le développement d'applications 'Threaded' qui utilisent le har

    Téléchargez gratuitement l'ebook sur le développement d'applications ?Threaded' Les logiciels de développement Intel® Parallel Studio accélèrent le développement d'applications ?Threaded' qui utilisent le hardware des utilisateurs finaux, depuis le ?'supercomputer'' jusqu'à l'ordinateur portable ou les mobiles. Optimisez la performance de votre application sur architecture Intel® et obtenez plus des derniers processeurs multi-coeurs d'Intel®. Depuis la manière dont les produits fonctionnent ensemble jusqu'à leurs jeux de fonctionnalités uniques, le Threading est maintenant plus facile et plus viable que jamais. Les outils sont optimisés donc les novices peuvent facilement se former et les développeurs expérimentés peuvent aisément ...

    Read the article

  • The Birth and Life of a Disk Galaxy [Video]

    - by Jason Fitzpatrick
    In this video, rendered over a million CPU hours by the Pleiades supercomputer at NASA’s Ames Research Center, we see the birth and life of a massive disk galaxy. Computer Model Shows a Disk Galaxy’s Life History [via Geeks Are Sexy] HTG Explains: Why It’s Good That Your Computer’s RAM Is Full 10 Awesome Improvements For Desktop Users in Windows 8 How To Play DVDs on Windows 8

    Read the article

  • A new number one

    - by nospam(at)example.com (Joerg Moellenkamp)
    The Top500 supercomputer list has a new number one: The K Computer, built by Fujitsu, currently combines 68544 SPARC64 VIIIfx CPUs, each with eight cores, for a total of 548,352 cores?almost twice as many as any other system in the TOP500. The K Computer is also more powerful than the next five systems on the list combined.Interestingly this system runs under Linux. And it uses tofu as its interconnect

    Read the article

  • Azure Futures - Distributed Computing and Number Crunching

    - by JoshReuben
    "the biggest Azure customers today are the ones using HPC on-premises at the current time" - http://www.zdnet.com/blog/microsoft/windows-azure-futures-turning-the-cloud-into-a-supercomputer/8592?utm_source=feedburner&utm_medium=feed&utm_campaign=Feed%3A+zdnet%2Fmicrosoft+%28ZDNet+All+About+Microsoft%29&utm_content=Google+Reader   Orleans Framework for cloud computing - http://research.microsoft.com/en-us/projects/orleans     HPC on Azure - http://www.zdnet.com/blog/microsoft/microsoft-finalizes-its-latest-supercomputing-operating-system-release/7414   Dryad is Microsoft’s competitor to Google MapReduce and Apache Hadoop  - http://www.zdnet.com/blog/microsoft/microsoft-takes-a-step-toward-commercializing-its-dryad-distributed-computing-technologies/8255?tag=mantle_skin;content   SQL Server Analysis Services DataMining in the cloud - http://www.sqlmag.com/article/reporting2/azure-data-mining-in-the-cloud.aspx

    Read the article

  • Intro to GPU programming

    - by Adam Davis
    Everyone has this huge massively parallelized supercomputer on their desktop in the form of a graphics card GPU. What is the "hello world" equivalent of the GPU community? What do I do, where do I go, to get started programming the GPU for the major GPU vendors? -Adam

    Read the article

  • writing large excel spreadsheets

    - by pstanton
    has anybody found a library that works well with large spreadsheets? I've tried apache's POI but it fails miserably working with large files - both reading and writing. It uses massive amounts of memory leaving you needing a supercomputer to parse or create a 20+mb spreadsheet. Surely there is a more memory efficient way and someone has written it?!

    Read the article

  • Javascript: Machine Constants Applicable?

    - by DavidB2013
    I write numerical routines for students of science and engineering (although they are freely available for use by anybody else as well) and am wondering how to properly use machine constants in a JavaScript program, or if they are even applicable. For example, say I am writing a program in C++ that numerically computes the roots of the following equation: exp(-0.7x) + sin(3x) - 1.2x + 0.3546 = 0 A root-finding routine should be able to compute roots to within the machine epsilon. In C++, this value is specified by the language: DBL_EPSILON. C++ also specifies the smallest and largest values that can be held by a float or double variable. However, how does this convert to JavaScript? Since a Javascript program runs in a web browser, and I don't know what kind of computer will run the program, and JavaScript does not have corresponding predefined values for these quantities, how can I implement my own version of these constants so that my programs compute results to as much accuracy as allowed on the computer running the web browser? My first draft is to simply copy over the literal constants from C++: FLT_MIN: 1.17549435082229e-038 FLT_MAX: 3.40282346638529e+038 DBL_EPSILON: 2.2204460492503131e-16 I am also willing to write small code blocks that could compute these values for each machine on which the program is run. That way, a supercomputer might compute results to a higher accuracy than an old, low-level, PC. BUT, I don't know if such a routine would actually reach the computer, in which case, I would be wasting my time. Anybody here know how to compute and use (in Javascript) values that correspond to machine constants in a compiled language? Is it worth my time to write small programs in Javascript that compute DBL_EPSILON, FLT_MIN, FLT_MIN, etc. for use in numerical routines? Or am I better off simply assigning literal constants that come straight from C++ on a standard Windows PC?

    Read the article

  • A way to auto cycle (close) through all screen sessions

    - by JBWhitmore
    I frequently use screen when I log into the interactive nodes to a supercomputer that I have access to -- and I often run things and move on. There are about 20 separate nodes that I can log into; and if I check any one of them I'll have something like 4 detached sessions. Each of those sessions will have maybe 5 screen sessions within that. Is there a quick way to cycle through all of these and close them down if they are not running any processes? My current process is to screen -ls and then screen -r #### then type exit until I'm back to the base screen.

    Read the article

  • Somewhat powerful server needed for computationally expensive stuff

    - by Dane Larsen
    So here's my problem. My Dad runs a company that does some rather computationally expensive stuff. This is not supercomputer level stuff, but it does take several hours to run the average job on his Core i7 desktop. He asked me to look into a way to have his customers use the code on an hourly basis, namely via a server. Ideally he'd be able to buy a box for about $1000, and hook it right up to our home connection. Unfortunately, the data that needs to be both sent and received is on the order of several hundred megs. We live in a rural area, and the fastest connection offered is 1.5Mbit/s. Download. It's like .3Mbit/s upload. Not workable. What are the options for this kind of thing? Ideally, we'd have about 2GB of ram, 300-500GB of storage, and a nice dual core, and it has to run some flavor of Linux. Any suggestions? Thanks in advance EDIT: Also, ideally the monthly price would be < $100 per month.

    Read the article

  • About the K computer

    - by nospam(at)example.com (Joerg Moellenkamp)
    Okay ? after getting yet another mail because of the new #1 on the Top500 list, I want to add some comments from my side: Yes, the system is using SPARC processor. And that is great news for a SPARC fan like me. It is using the SPARC VIIIfx processor from Fujitsu clocked at 2 GHz. No, it isn't the only one. Most people are saying there are two in the Top500 list using SPARC (#77 JAXA and #1 K) but in fact there are three. The Tianhe-1 (#2 on the Top500 list) super computer contains 2048 Galaxy "FT-1000" 1 GHz 8-core processors. Don't know it? The FeiTeng-1000 ? this proc is a 8 core, 8 threads per core, 1 ghz processor made in China. And it's SPARC based. By the way ? this sounds really familiar to me ? perhaps the people just took the opensourced UltraSPARC-T2 design, because some of the parameters sound just to similar. However it looks like that Tianhe-1 is using the SPARCs as input nodes and not as compute notes. No, I don't see it as the next M-series processor. Simple reason: You can't create SMP systems out of them ? it simply hasn't the functionality to do so. Even when there are multiple CPUs on a single board, they are not connected like an SMP/NUMA machine to a shared memory machine ? they are connected with the cluster interconnect (in this case the Tofu interconnect) and work like a large cluster. Yes, it has a lot of oomph in Linpack ? however I assume a lot came from the extensions to the SPARCv9 standard. No, Linpack has no relevance for any commercial workload ? Linpack is such a special load, that even some HPC people are arguing that it isn't really a good benchmark for HPC. It's embarrassingly parallel, it can work with relatively small interconnects compared to the interconnects in SMP systems (however we get in spheres SMP interconnects where a few years ago). Amdahl isn't hitting that hard when running Linpack. Yes, it's a good move to use SPARC. At some time in the last 10 years, there was an interesting twist in perception: SPARC was considered as proprietary architecture and x86 was the open architecture. However it's vice versa ? try to create a x86 clone and you have a lot of intellectual property problems, create a SPARC clone and you have to spend 100 bucks or so to get the specification from the SPARC Foundation and develop your own SPARC processor. Fujitsu is doing this for a long time now. So they had their own processor, their own know-how. So why was SPARC a good choice? Well ? essentially Fujitsu can do what they want with their core as it is their core, for example adding the extensions to the SPARCv9 chipset ? getting Intel to create extensions to x86 to help you with your product is a little bit harder. So Fujitsu could do they needed to do with their processor in order to create such a supercomputer. No, the K is really using no FPGA or GPU as accelerators. The K is really using the CPU at doing this job. Yes, it has a significantly enhanced FPU capable to execute 8 instructions in parallel. No, it doesn't run Solaris. Yes, it uses Linux. No, it doesn't hurt me ... as my colleague Roland Rambau (he knows a lot about HPC) said once to me ... it doesn't matter which OS is staying out of the way of the workload in HPC.

    Read the article

  • How to get the best LINPACK result and conquer the Top500?

    - by knweiss
    Given a large Linux HPC cluster with hundreds/thousands of nodes. What are your best practices to get the best possible LINPACK benchmark (HPL) result to submit for the Top500 supercomputer list? To give you an idea what kind of answers I would appreciate here are some sub-questions (with links): How to you tune the parameters (N, NB, P, Q, memory-alignment, etc) for the HPL.dat file (without spending too much time trying each possible permutation - esp with large problem sizes N)? Are there any Top500 submission rules to be aware of? What is allowed, what isn't? Which MPI product, which version? Does it make a difference? Any special host order in your MPI machine file? Do you use CPU pinning? How to you configure your interconnect? Which interconnect? Which BLAS package do you use for which CPU model? (Intel MKL, AMD ACML, GotoBLAS2, etc.) How do you prepare for the big run (on all nodes)? Start with small runs on a subset of nodes and then scale up? Is it really necessary to run LINPACK with a big run on all of the nodes (or is extrapolation allowed)? How do you optimize for the latest Intel/AMD CPUs? Hyperthreading? NUMA? Is it worth it to recompile the software stack or do you use precompiled binaries? Which settings? Which compiler optimizations, which compiler? (What about profile-based compilation?) How to get the best result given only a limited amount of time to do the benchmark run? (You can block a huge cluster forever) How do you prepare the individual nodes (stopping system daemons, freeing memory, etc)? How do you deal with hardware faults (ruining a huge run)? Are there any must-read documents or websites about this topic? E.g. I would love to hear about some background stories of some of the current Top500 systems and how they did their LINPACK benchmark. I deliberately don't want to mention concrete hardware details or discuss hardware recommendations because I don't want to limit the answers. However, feel free to mention hints e.g. for specific CPU models.

    Read the article

  • Windows Azure Myths

    - by BuckWoody
    Windows Azure is part of the Microsoft "stack" - the suite of software and services we offer. Because we have so many products in almost every part of technology, it's hard to know everything about all parts of what we do - even for those of us who work here. So it's no surprise that some folks are not as familiar with Windows and SQL Azure as they are, say Windows Server or XBox. As I chat with folks about a solution for a business or organization need, I put Windows Azure into the mix. I always start off with "What do you already know about Windows Azure?" so that I don't bore folks with information they already have. I some cases they've checked out the product ahead of time and have specific questions, in others they aren't as familiar, and in still others there is a fair amount of mis-information. Sometimes that's because of a marketing failure, sometimes it's hearsay, and somtetimes it's active misinformation. I thought I might lay out a few of these misconceptions. As always - do your fact-checking! Never take anyone's word alone (including mine) as gospel. Make sure you educate yourself on your options. Your company or your clients depend on you to have the right information on IT, so make sure you live up to that. Myth 1: Nobody uses Windows Azure It's true that we don't give out numbers on the amount of clients on Windows and SQL Azure. But lots of folks are here - companies you may have heard of like Boeing, NASA, Fujitsu, The City of London, Nuedesic, and many others. I deal with firms small and large that use Windows Azure for mission-critical applications, sometimes totally on Windows and/or SQL Azure, sometimes in conjunction with an on-premises system, sometimes for only a specific component in Windows Azure like storage. The interesting thing is that many sites you visit have a Windows Azure component, or are running on Windows Azure. They just don't announce it. Just like the other cloud providers, the companies have asked to be completely branded themselves - they don't want you to be aware or care that they are on Windows Azure. Sometimes that's for security, other times it's for different reasons. It's just like the web sites you visit. For the most part, they don't advertise which OS or Web Server they use. It really just shouldn't matter. The point is that they just use what works to solve a given problem. Check out a few public case studies here: https://www.windowsazure.com/en-us/home/case-studies/ Myth 2: It's only for Microsoft stuff - can't use Open Source This is the one I face the most, and am the most dismayed by. We work just fine with many open source products, including Java, NodeJS, PHP, Ruby, Python, Hadoop, and many other languages and applications. You can quickly deploy a Wordpress, Umbraco and other "kits". We have software development kits (SDK's) for iPhones, iPads, Android, Windows phones and more. We have an SDK to work with FaceBook and other social networks. In short, we play well with others. More on the languages and runtimes we support here: https://www.windowsazure.com/en-us/develop/overview/ More on the SDK's here: http://www.wadewegner.com/2011/05/windows-azure-toolkit-for-ios/, http://www.wadewegner.com/2011/08/windows-azure-toolkits-for-devices-now-with-android/, http://azuretoolkit.codeplex.com/ Myth 3: Microsoft expects me to switch everything to "the cloud" No, we don't. That would be disasterous, unless the only things you run in your company uses works perfectly in Azure. Use Windows Azure  - or any cloud for that matter - where it works. Whenever I talk to companies, I focus on two things: Something that is broken and needs to be re-architected Something you want to do that is new If something is broken, and you need new tools to scale, extend, add capacity dynamically and so on, then you can consider using Windows or SQL Azure. It can help solve problems that you have, or it may include a component you don't want to write or architect yourself. Sometimes you want to do something new, like extend your company's offerings to mobile phones, to the web, or to a social network. More info on where it works here: http://blogs.msdn.com/b/buckwoody/archive/2011/01/18/windows-azure-and-sql-azure-use-cases.aspx Myth 4: I have to write code to use Windows and SQL Azure If Windows Azure is a PaaS - a Platform as a Service - then don't you have to write code to use it? Nope. Windows and SQL Azure are made up of various components. Some of those components allow you to write and deploy code (like Compute) and others don't. We have lots of customers using Windows Azure storage as a backup, to securely share files instead of using DropBox, to distribute videos or code or firmware, and more. Others use our High Performance Computing (HPC) offering to rent a supercomputer when they need one. You can even throw workloads at that using Excel! In addition there are lots of other components in Windows Azure you can use, from the Windows Azure Media Services to others. More here: https://www.windowsazure.com/en-us/home/scenarios/saas/ Myth 5: Windows Azure is just another form of "vendor lock-in" Windows Azure uses .NET, OSS languages and standard interfaces for the code. Sure, you're not going to take the code line-for-line and run it on a mainframe, but it's standard code that you write, and can port to something else. And the data is yours - you can bring it back whever you want. It's either in text or binary form, that you have complete control over. There are no licenses - you can "pay as you go", and when you're done, you can leave the service and take all your code, data and IP with you.   So go out there, read up, try it. Use it where it works. And don't believe everything you hear - sometimes the Internet doesn't get it all correct. :)

    Read the article

  • Windows Azure Myths

    - by BuckWoody
    Windows Azure is part of the Microsoft "stack" - the suite of software and services we offer. Because we have so many products in almost every part of technology, it's hard to know everything about all parts of what we do - even for those of us who work here. So it's no surprise that some folks are not as familiar with Windows and SQL Azure as they are, say Windows Server or XBox. As I chat with folks about a solution for a business or organization need, I put Windows Azure into the mix. I always start off with "What do you already know about Windows Azure?" so that I don't bore folks with information they already have. I some cases they've checked out the product ahead of time and have specific questions, in others they aren't as familiar, and in still others there is a fair amount of mis-information. Sometimes that's because of a marketing failure, sometimes it's hearsay, and somtetimes it's active misinformation. I thought I might lay out a few of these misconceptions. As always - do your fact-checking! Never take anyone's word alone (including mine) as gospel. Make sure you educate yourself on your options. Your company or your clients depend on you to have the right information on IT, so make sure you live up to that. Myth 1: Nobody uses Windows Azure It's true that we don't give out numbers on the amount of clients on Windows and SQL Azure. But lots of folks are here - companies you may have heard of like Boeing, NASA, Fujitsu, The City of London, Nuedesic, and many others. I deal with firms small and large that use Windows Azure for mission-critical applications, sometimes totally on Windows and/or SQL Azure, sometimes in conjunction with an on-premises system, sometimes for only a specific component in Windows Azure like storage. The interesting thing is that many sites you visit have a Windows Azure component, or are running on Windows Azure. They just don't announce it. Just like the other cloud providers, the companies have asked to be completely branded themselves - they don't want you to be aware or care that they are on Windows Azure. Sometimes that's for security, other times it's for different reasons. It's just like the web sites you visit. For the most part, they don't advertise which OS or Web Server they use. It really just shouldn't matter. The point is that they just use what works to solve a given problem. Check out a few public case studies here: https://www.windowsazure.com/en-us/home/case-studies/ Myth 2: It's only for Microsoft stuff - can't use Open Source This is the one I face the most, and am the most dismayed by. We work just fine with many open source products, including Java, NodeJS, PHP, Ruby, Python, Hadoop, and many other languages and applications. You can quickly deploy a Wordpress, Umbraco and other "kits". We have software development kits (SDK's) for iPhones, iPads, Android, Windows phones and more. We have an SDK to work with FaceBook and other social networks. In short, we play well with others. More on the languages and runtimes we support here: https://www.windowsazure.com/en-us/develop/overview/ More on the SDK's here: http://www.wadewegner.com/2011/05/windows-azure-toolkit-for-ios/, http://www.wadewegner.com/2011/08/windows-azure-toolkits-for-devices-now-with-android/, http://azuretoolkit.codeplex.com/ Myth 3: Microsoft expects me to switch everything to "the cloud" No, we don't. That would be disasterous, unless the only things you run in your company uses works perfectly in Azure. Use Windows Azure  - or any cloud for that matter - where it works. Whenever I talk to companies, I focus on two things: Something that is broken and needs to be re-architected Something you want to do that is new If something is broken, and you need new tools to scale, extend, add capacity dynamically and so on, then you can consider using Windows or SQL Azure. It can help solve problems that you have, or it may include a component you don't want to write or architect yourself. Sometimes you want to do something new, like extend your company's offerings to mobile phones, to the web, or to a social network. More info on where it works here: http://blogs.msdn.com/b/buckwoody/archive/2011/01/18/windows-azure-and-sql-azure-use-cases.aspx Myth 4: I have to write code to use Windows and SQL Azure If Windows Azure is a PaaS - a Platform as a Service - then don't you have to write code to use it? Nope. Windows and SQL Azure are made up of various components. Some of those components allow you to write and deploy code (like Compute) and others don't. We have lots of customers using Windows Azure storage as a backup, to securely share files instead of using DropBox, to distribute videos or code or firmware, and more. Others use our High Performance Computing (HPC) offering to rent a supercomputer when they need one. You can even throw workloads at that using Excel! In addition there are lots of other components in Windows Azure you can use, from the Windows Azure Media Services to others. More here: https://www.windowsazure.com/en-us/home/scenarios/saas/ Myth 5: Windows Azure is just another form of "vendor lock-in" Windows Azure uses .NET, OSS languages and standard interfaces for the code. Sure, you're not going to take the code line-for-line and run it on a mainframe, but it's standard code that you write, and can port to something else. And the data is yours - you can bring it back whever you want. It's either in text or binary form, that you have complete control over. There are no licenses - you can "pay as you go", and when you're done, you can leave the service and take all your code, data and IP with you.   So go out there, read up, try it. Use it where it works. And don't believe everything you hear - sometimes the Internet doesn't get it all correct. :)

    Read the article

  • How John Got 15x Improvement Without Really Trying

    - by rchrd
    The following article was published on a Sun Microsystems website a number of years ago by John Feo. It is still useful and worth preserving. So I'm republishing it here.  How I Got 15x Improvement Without Really Trying John Feo, Sun Microsystems Taking ten "personal" program codes used in scientific and engineering research, the author was able to get from 2 to 15 times performance improvement easily by applying some simple general optimization techniques. Introduction Scientific research based on computer simulation depends on the simulation for advancement. The research can advance only as fast as the computational codes can execute. The codes' efficiency determines both the rate and quality of results. In the same amount of time, a faster program can generate more results and can carry out a more detailed simulation of physical phenomena than a slower program. Highly optimized programs help science advance quickly and insure that monies supporting scientific research are used as effectively as possible. Scientific computer codes divide into three broad categories: ISV, community, and personal. ISV codes are large, mature production codes developed and sold commercially. The codes improve slowly over time both in methods and capabilities, and they are well tuned for most vendor platforms. Since the codes are mature and complex, there are few opportunities to improve their performance solely through code optimization. Improvements of 10% to 15% are typical. Examples of ISV codes are DYNA3D, Gaussian, and Nastran. Community codes are non-commercial production codes used by a particular research field. Generally, they are developed and distributed by a single academic or research institution with assistance from the community. Most users just run the codes, but some develop new methods and extensions that feed back into the general release. The codes are available on most vendor platforms. Since these codes are younger than ISV codes, there are more opportunities to optimize the source code. Improvements of 50% are not unusual. Examples of community codes are AMBER, CHARM, BLAST, and FASTA. Personal codes are those written by single users or small research groups for their own use. These codes are not distributed, but may be passed from professor-to-student or student-to-student over several years. They form the primordial ocean of applications from which community and ISV codes emerge. Government research grants pay for the development of most personal codes. This paper reports on the nature and performance of this class of codes. Over the last year, I have looked at over two dozen personal codes from more than a dozen research institutions. The codes cover a variety of scientific fields, including astronomy, atmospheric sciences, bioinformatics, biology, chemistry, geology, and physics. The sources range from a few hundred lines to more than ten thousand lines, and are written in Fortran, Fortran 90, C, and C++. For the most part, the codes are modular, documented, and written in a clear, straightforward manner. They do not use complex language features, advanced data structures, programming tricks, or libraries. I had little trouble understanding what the codes did or how data structures were used. Most came with a makefile. Surprisingly, only one of the applications is parallel. All developers have access to parallel machines, so availability is not an issue. Several tried to parallelize their applications, but stopped after encountering difficulties. Lack of education and a perception that parallelism is difficult prevented most from trying. I parallelized several of the codes using OpenMP, and did not judge any of the codes as difficult to parallelize. Even more surprising than the lack of parallelism is the inefficiency of the codes. I was able to get large improvements in performance in a matter of a few days applying simple optimization techniques. Table 1 lists ten representative codes [names and affiliation are omitted to preserve anonymity]. Improvements on one processor range from 2x to 15.5x with a simple average of 4.75x. I did not use sophisticated performance tools or drill deep into the program's execution character as one would do when tuning ISV or community codes. Using only a profiler and source line timers, I identified inefficient sections of code and improved their performance by inspection. The changes were at a high level. I am sure there is another factor of 2 or 3 in each code, and more if the codes are parallelized. The study’s results show that personal scientific codes are running many times slower than they should and that the problem is pervasive. Computational scientists are not sloppy programmers; however, few are trained in the art of computer programming or code optimization. I found that most have a working knowledge of some programming language and standard software engineering practices; but they do not know, or think about, how to make their programs run faster. They simply do not know the standard techniques used to make codes run faster. In fact, they do not even perceive that such techniques exist. The case studies described in this paper show that applying simple, well known techniques can significantly increase the performance of personal codes. It is important that the scientific community and the Government agencies that support scientific research find ways to better educate academic scientific programmers. The inefficiency of their codes is so bad that it is retarding both the quality and progress of scientific research. # cacheperformance redundantoperations loopstructures performanceimprovement 1 x x 15.5 2 x 2.8 3 x x 2.5 4 x 2.1 5 x x 2.0 6 x 5.0 7 x 5.8 8 x 6.3 9 2.2 10 x x 3.3 Table 1 — Area of improvement and performance gains of 10 codes The remainder of the paper is organized as follows: sections 2, 3, and 4 discuss the three most common sources of inefficiencies in the codes studied. These are cache performance, redundant operations, and loop structures. Each section includes several examples. The last section summaries the work and suggests a possible solution to the issues raised. Optimizing cache performance Commodity microprocessor systems use caches to increase memory bandwidth and reduce memory latencies. Typical latencies from processor to L1, L2, local, and remote memory are 3, 10, 50, and 200 cycles, respectively. Moreover, bandwidth falls off dramatically as memory distances increase. Programs that do not use cache effectively run many times slower than programs that do. When optimizing for cache, the biggest performance gains are achieved by accessing data in cache order and reusing data to amortize the overhead of cache misses. Secondary considerations are prefetching, associativity, and replacement; however, the understanding and analysis required to optimize for the latter are probably beyond the capabilities of the non-expert. Much can be gained simply by accessing data in the correct order and maximizing data reuse. 6 out of the 10 codes studied here benefited from such high level optimizations. Array Accesses The most important cache optimization is the most basic: accessing Fortran array elements in column order and C array elements in row order. Four of the ten codes—1, 2, 4, and 10—got it wrong. Compilers will restructure nested loops to optimize cache performance, but may not do so if the loop structure is too complex, or the loop body includes conditionals, complex addressing, or function calls. In code 1, the compiler failed to invert a key loop because of complex addressing do I = 0, 1010, delta_x IM = I - delta_x IP = I + delta_x do J = 5, 995, delta_x JM = J - delta_x JP = J + delta_x T1 = CA1(IP, J) + CA1(I, JP) T2 = CA1(IM, J) + CA1(I, JM) S1 = T1 + T2 - 4 * CA1(I, J) CA(I, J) = CA1(I, J) + D * S1 end do end do In code 2, the culprit is conditionals do I = 1, N do J = 1, N If (IFLAG(I,J) .EQ. 0) then T1 = Value(I, J-1) T2 = Value(I-1, J) T3 = Value(I, J) T4 = Value(I+1, J) T5 = Value(I, J+1) Value(I,J) = 0.25 * (T1 + T2 + T5 + T4) Delta = ABS(T3 - Value(I,J)) If (Delta .GT. MaxDelta) MaxDelta = Delta endif enddo enddo I fixed both programs by inverting the loops by hand. Code 10 has three-dimensional arrays and triply nested loops. The structure of the most computationally intensive loops is too complex to invert automatically or by hand. The only practical solution is to transpose the arrays so that the dimension accessed by the innermost loop is in cache order. The arrays can be transposed at construction or prior to entering a computationally intensive section of code. The former requires all array references to be modified, while the latter is cost effective only if the cost of the transpose is amortized over many accesses. I used the second approach to optimize code 10. Code 5 has four-dimensional arrays and loops are nested four deep. For all of the reasons cited above the compiler is not able to restructure three key loops. Assume C arrays and let the four dimensions of the arrays be i, j, k, and l. In the original code, the index structure of the three loops is L1: for i L2: for i L3: for i for l for l for j for k for j for k for j for k for l So only L3 accesses array elements in cache order. L1 is a very complex loop—much too complex to invert. I brought the loop into cache alignment by transposing the second and fourth dimensions of the arrays. Since the code uses a macro to compute all array indexes, I effected the transpose at construction and changed the macro appropriately. The dimensions of the new arrays are now: i, l, k, and j. L3 is a simple loop and easily inverted. L2 has a loop-carried scalar dependence in k. By promoting the scalar name that carries the dependence to an array, I was able to invert the third and fourth subloops aligning the loop with cache. Code 5 is by far the most difficult of the four codes to optimize for array accesses; but the knowledge required to fix the problems is no more than that required for the other codes. I would judge this code at the limits of, but not beyond, the capabilities of appropriately trained computational scientists. Array Strides When a cache miss occurs, a line (64 bytes) rather than just one word is loaded into the cache. If data is accessed stride 1, than the cost of the miss is amortized over 8 words. Any stride other than one reduces the cost savings. Two of the ten codes studied suffered from non-unit strides. The codes represent two important classes of "strided" codes. Code 1 employs a multi-grid algorithm to reduce time to convergence. The grids are every tenth, fifth, second, and unit element. Since time to convergence is inversely proportional to the distance between elements, coarse grids converge quickly providing good starting values for finer grids. The better starting values further reduce the time to convergence. The downside is that grids of every nth element, n > 1, introduce non-unit strides into the computation. In the original code, much of the savings of the multi-grid algorithm were lost due to this problem. I eliminated the problem by compressing (copying) coarse grids into continuous memory, and rewriting the computation as a function of the compressed grid. On convergence, I copied the final values of the compressed grid back to the original grid. The savings gained from unit stride access of the compressed grid more than paid for the cost of copying. Using compressed grids, the loop from code 1 included in the previous section becomes do j = 1, GZ do i = 1, GZ T1 = CA(i+0, j-1) + CA(i-1, j+0) T4 = CA1(i+1, j+0) + CA1(i+0, j+1) S1 = T1 + T4 - 4 * CA1(i+0, j+0) CA(i+0, j+0) = CA1(i+0, j+0) + DD * S1 enddo enddo where CA and CA1 are compressed arrays of size GZ. Code 7 traverses a list of objects selecting objects for later processing. The labels of the selected objects are stored in an array. The selection step has unit stride, but the processing steps have irregular stride. A fix is to save the parameters of the selected objects in temporary arrays as they are selected, and pass the temporary arrays to the processing functions. The fix is practical if the same parameters are used in selection as in processing, or if processing comprises a series of distinct steps which use overlapping subsets of the parameters. Both conditions are true for code 7, so I achieved significant improvement by copying parameters to temporary arrays during selection. Data reuse In the previous sections, we optimized for spatial locality. It is also important to optimize for temporal locality. Once read, a datum should be used as much as possible before it is forced from cache. Loop fusion and loop unrolling are two techniques that increase temporal locality. Unfortunately, both techniques increase register pressure—as loop bodies become larger, the number of registers required to hold temporary values grows. Once register spilling occurs, any gains evaporate quickly. For multiprocessors with small register sets or small caches, the sweet spot can be very small. In the ten codes presented here, I found no opportunities for loop fusion and only two opportunities for loop unrolling (codes 1 and 3). In code 1, unrolling the outer and inner loop one iteration increases the number of result values computed by the loop body from 1 to 4, do J = 1, GZ-2, 2 do I = 1, GZ-2, 2 T1 = CA1(i+0, j-1) + CA1(i-1, j+0) T2 = CA1(i+1, j-1) + CA1(i+0, j+0) T3 = CA1(i+0, j+0) + CA1(i-1, j+1) T4 = CA1(i+1, j+0) + CA1(i+0, j+1) T5 = CA1(i+2, j+0) + CA1(i+1, j+1) T6 = CA1(i+1, j+1) + CA1(i+0, j+2) T7 = CA1(i+2, j+1) + CA1(i+1, j+2) S1 = T1 + T4 - 4 * CA1(i+0, j+0) S2 = T2 + T5 - 4 * CA1(i+1, j+0) S3 = T3 + T6 - 4 * CA1(i+0, j+1) S4 = T4 + T7 - 4 * CA1(i+1, j+1) CA(i+0, j+0) = CA1(i+0, j+0) + DD * S1 CA(i+1, j+0) = CA1(i+1, j+0) + DD * S2 CA(i+0, j+1) = CA1(i+0, j+1) + DD * S3 CA(i+1, j+1) = CA1(i+1, j+1) + DD * S4 enddo enddo The loop body executes 12 reads, whereas as the rolled loop shown in the previous section executes 20 reads to compute the same four values. In code 3, two loops are unrolled 8 times and one loop is unrolled 4 times. Here is the before for (k = 0; k < NK[u]; k++) { sum = 0.0; for (y = 0; y < NY; y++) { sum += W[y][u][k] * delta[y]; } backprop[i++]=sum; } and after code for (k = 0; k < KK - 8; k+=8) { sum0 = 0.0; sum1 = 0.0; sum2 = 0.0; sum3 = 0.0; sum4 = 0.0; sum5 = 0.0; sum6 = 0.0; sum7 = 0.0; for (y = 0; y < NY; y++) { sum0 += W[y][0][k+0] * delta[y]; sum1 += W[y][0][k+1] * delta[y]; sum2 += W[y][0][k+2] * delta[y]; sum3 += W[y][0][k+3] * delta[y]; sum4 += W[y][0][k+4] * delta[y]; sum5 += W[y][0][k+5] * delta[y]; sum6 += W[y][0][k+6] * delta[y]; sum7 += W[y][0][k+7] * delta[y]; } backprop[k+0] = sum0; backprop[k+1] = sum1; backprop[k+2] = sum2; backprop[k+3] = sum3; backprop[k+4] = sum4; backprop[k+5] = sum5; backprop[k+6] = sum6; backprop[k+7] = sum7; } for one of the loops unrolled 8 times. Optimizing for temporal locality is the most difficult optimization considered in this paper. The concepts are not difficult, but the sweet spot is small. Identifying where the program can benefit from loop unrolling or loop fusion is not trivial. Moreover, it takes some effort to get it right. Still, educating scientific programmers about temporal locality and teaching them how to optimize for it will pay dividends. Reducing instruction count Execution time is a function of instruction count. Reduce the count and you usually reduce the time. The best solution is to use a more efficient algorithm; that is, an algorithm whose order of complexity is smaller, that converges quicker, or is more accurate. Optimizing source code without changing the algorithm yields smaller, but still significant, gains. This paper considers only the latter because the intent is to study how much better codes can run if written by programmers schooled in basic code optimization techniques. The ten codes studied benefited from three types of "instruction reducing" optimizations. The two most prevalent were hoisting invariant memory and data operations out of inner loops. The third was eliminating unnecessary data copying. The nature of these inefficiencies is language dependent. Memory operations The semantics of C make it difficult for the compiler to determine all the invariant memory operations in a loop. The problem is particularly acute for loops in functions since the compiler may not know the values of the function's parameters at every call site when compiling the function. Most compilers support pragmas to help resolve ambiguities; however, these pragmas are not comprehensive and there is no standard syntax. To guarantee that invariant memory operations are not executed repetitively, the user has little choice but to hoist the operations by hand. The problem is not as severe in Fortran programs because in the absence of equivalence statements, it is a violation of the language's semantics for two names to share memory. Codes 3 and 5 are C programs. In both cases, the compiler did not hoist all invariant memory operations from inner loops. Consider the following loop from code 3 for (y = 0; y < NY; y++) { i = 0; for (u = 0; u < NU; u++) { for (k = 0; k < NK[u]; k++) { dW[y][u][k] += delta[y] * I1[i++]; } } } Since dW[y][u] can point to the same memory space as delta for one or more values of y and u, assignment to dW[y][u][k] may change the value of delta[y]. In reality, dW and delta do not overlap in memory, so I rewrote the loop as for (y = 0; y < NY; y++) { i = 0; Dy = delta[y]; for (u = 0; u < NU; u++) { for (k = 0; k < NK[u]; k++) { dW[y][u][k] += Dy * I1[i++]; } } } Failure to hoist invariant memory operations may be due to complex address calculations. If the compiler can not determine that the address calculation is invariant, then it can hoist neither the calculation nor the associated memory operations. As noted above, code 5 uses a macro to address four-dimensional arrays #define MAT4D(a,q,i,j,k) (double *)((a)->data + (q)*(a)->strides[0] + (i)*(a)->strides[3] + (j)*(a)->strides[2] + (k)*(a)->strides[1]) The macro is too complex for the compiler to understand and so, it does not identify any subexpressions as loop invariant. The simplest way to eliminate the address calculation from the innermost loop (over i) is to define a0 = MAT4D(a,q,0,j,k) before the loop and then replace all instances of *MAT4D(a,q,i,j,k) in the loop with a0[i] A similar problem appears in code 6, a Fortran program. The key loop in this program is do n1 = 1, nh nx1 = (n1 - 1) / nz + 1 nz1 = n1 - nz * (nx1 - 1) do n2 = 1, nh nx2 = (n2 - 1) / nz + 1 nz2 = n2 - nz * (nx2 - 1) ndx = nx2 - nx1 ndy = nz2 - nz1 gxx = grn(1,ndx,ndy) gyy = grn(2,ndx,ndy) gxy = grn(3,ndx,ndy) balance(n1,1) = balance(n1,1) + (force(n2,1) * gxx + force(n2,2) * gxy) * h1 balance(n1,2) = balance(n1,2) + (force(n2,1) * gxy + force(n2,2) * gyy)*h1 end do end do The programmer has written this loop well—there are no loop invariant operations with respect to n1 and n2. However, the loop resides within an iterative loop over time and the index calculations are independent with respect to time. Trading space for time, I precomputed the index values prior to the entering the time loop and stored the values in two arrays. I then replaced the index calculations with reads of the arrays. Data operations Ways to reduce data operations can appear in many forms. Implementing a more efficient algorithm produces the biggest gains. The closest I came to an algorithm change was in code 4. This code computes the inner product of K-vectors A(i) and B(j), 0 = i < N, 0 = j < M, for most values of i and j. Since the program computes most of the NM possible inner products, it is more efficient to compute all the inner products in one triply-nested loop rather than one at a time when needed. The savings accrue from reading A(i) once for all B(j) vectors and from loop unrolling. for (i = 0; i < N; i+=8) { for (j = 0; j < M; j++) { sum0 = 0.0; sum1 = 0.0; sum2 = 0.0; sum3 = 0.0; sum4 = 0.0; sum5 = 0.0; sum6 = 0.0; sum7 = 0.0; for (k = 0; k < K; k++) { sum0 += A[i+0][k] * B[j][k]; sum1 += A[i+1][k] * B[j][k]; sum2 += A[i+2][k] * B[j][k]; sum3 += A[i+3][k] * B[j][k]; sum4 += A[i+4][k] * B[j][k]; sum5 += A[i+5][k] * B[j][k]; sum6 += A[i+6][k] * B[j][k]; sum7 += A[i+7][k] * B[j][k]; } C[i+0][j] = sum0; C[i+1][j] = sum1; C[i+2][j] = sum2; C[i+3][j] = sum3; C[i+4][j] = sum4; C[i+5][j] = sum5; C[i+6][j] = sum6; C[i+7][j] = sum7; }} This change requires knowledge of a typical run; i.e., that most inner products are computed. The reasons for the change, however, derive from basic optimization concepts. It is the type of change easily made at development time by a knowledgeable programmer. In code 5, we have the data version of the index optimization in code 6. Here a very expensive computation is a function of the loop indices and so cannot be hoisted out of the loop; however, the computation is invariant with respect to an outer iterative loop over time. We can compute its value for each iteration of the computation loop prior to entering the time loop and save the values in an array. The increase in memory required to store the values is small in comparison to the large savings in time. The main loop in Code 8 is doubly nested. The inner loop includes a series of guarded computations; some are a function of the inner loop index but not the outer loop index while others are a function of the outer loop index but not the inner loop index for (j = 0; j < N; j++) { for (i = 0; i < M; i++) { r = i * hrmax; R = A[j]; temp = (PRM[3] == 0.0) ? 1.0 : pow(r, PRM[3]); high = temp * kcoeff * B[j] * PRM[2] * PRM[4]; low = high * PRM[6] * PRM[6] / (1.0 + pow(PRM[4] * PRM[6], 2.0)); kap = (R > PRM[6]) ? high * R * R / (1.0 + pow(PRM[4]*r, 2.0) : low * pow(R/PRM[6], PRM[5]); < rest of loop omitted > }} Note that the value of temp is invariant to j. Thus, we can hoist the computation for temp out of the loop and save its values in an array. for (i = 0; i < M; i++) { r = i * hrmax; TEMP[i] = pow(r, PRM[3]); } [N.B. – the case for PRM[3] = 0 is omitted and will be reintroduced later.] We now hoist out of the inner loop the computations invariant to i. Since the conditional guarding the value of kap is invariant to i, it behooves us to hoist the computation out of the inner loop, thereby executing the guard once rather than M times. The final version of the code is for (j = 0; j < N; j++) { R = rig[j] / 1000.; tmp1 = kcoeff * par[2] * beta[j] * par[4]; tmp2 = 1.0 + (par[4] * par[4] * par[6] * par[6]); tmp3 = 1.0 + (par[4] * par[4] * R * R); tmp4 = par[6] * par[6] / tmp2; tmp5 = R * R / tmp3; tmp6 = pow(R / par[6], par[5]); if ((par[3] == 0.0) && (R > par[6])) { for (i = 1; i <= imax1; i++) KAP[i] = tmp1 * tmp5; } else if ((par[3] == 0.0) && (R <= par[6])) { for (i = 1; i <= imax1; i++) KAP[i] = tmp1 * tmp4 * tmp6; } else if ((par[3] != 0.0) && (R > par[6])) { for (i = 1; i <= imax1; i++) KAP[i] = tmp1 * TEMP[i] * tmp5; } else if ((par[3] != 0.0) && (R <= par[6])) { for (i = 1; i <= imax1; i++) KAP[i] = tmp1 * TEMP[i] * tmp4 * tmp6; } for (i = 0; i < M; i++) { kap = KAP[i]; r = i * hrmax; < rest of loop omitted > } } Maybe not the prettiest piece of code, but certainly much more efficient than the original loop, Copy operations Several programs unnecessarily copy data from one data structure to another. This problem occurs in both Fortran and C programs, although it manifests itself differently in the two languages. Code 1 declares two arrays—one for old values and one for new values. At the end of each iteration, the array of new values is copied to the array of old values to reset the data structures for the next iteration. This problem occurs in Fortran programs not included in this study and in both Fortran 77 and Fortran 90 code. Introducing pointers to the arrays and swapping pointer values is an obvious way to eliminate the copying; but pointers is not a feature that many Fortran programmers know well or are comfortable using. An easy solution not involving pointers is to extend the dimension of the value array by 1 and use the last dimension to differentiate between arrays at different times. For example, if the data space is N x N, declare the array (N, N, 2). Then store the problem’s initial values in (_, _, 2) and define the scalar names new = 2 and old = 1. At the start of each iteration, swap old and new to reset the arrays. The old–new copy problem did not appear in any C program. In programs that had new and old values, the code swapped pointers to reset data structures. Where unnecessary coping did occur is in structure assignment and parameter passing. Structures in C are handled much like scalars. Assignment causes the data space of the right-hand name to be copied to the data space of the left-hand name. Similarly, when a structure is passed to a function, the data space of the actual parameter is copied to the data space of the formal parameter. If the structure is large and the assignment or function call is in an inner loop, then copying costs can grow quite large. While none of the ten programs considered here manifested this problem, it did occur in programs not included in the study. A simple fix is always to refer to structures via pointers. Optimizing loop structures Since scientific programs spend almost all their time in loops, efficient loops are the key to good performance. Conditionals, function calls, little instruction level parallelism, and large numbers of temporary values make it difficult for the compiler to generate tightly packed, highly efficient code. Conditionals and function calls introduce jumps that disrupt code flow. Users should eliminate or isolate conditionls to their own loops as much as possible. Often logical expressions can be substituted for if-then-else statements. For example, code 2 includes the following snippet MaxDelta = 0.0 do J = 1, N do I = 1, M < code omitted > Delta = abs(OldValue ? NewValue) if (Delta > MaxDelta) MaxDelta = Delta enddo enddo if (MaxDelta .gt. 0.001) goto 200 Since the only use of MaxDelta is to control the jump to 200 and all that matters is whether or not it is greater than 0.001, I made MaxDelta a boolean and rewrote the snippet as MaxDelta = .false. do J = 1, N do I = 1, M < code omitted > Delta = abs(OldValue ? NewValue) MaxDelta = MaxDelta .or. (Delta .gt. 0.001) enddo enddo if (MaxDelta) goto 200 thereby, eliminating the conditional expression from the inner loop. A microprocessor can execute many instructions per instruction cycle. Typically, it can execute one or more memory, floating point, integer, and jump operations. To be executed simultaneously, the operations must be independent. Thick loops tend to have more instruction level parallelism than thin loops. Moreover, they reduce memory traffice by maximizing data reuse. Loop unrolling and loop fusion are two techniques to increase the size of loop bodies. Several of the codes studied benefitted from loop unrolling, but none benefitted from loop fusion. This observation is not too surpising since it is the general tendency of programmers to write thick loops. As loops become thicker, the number of temporary values grows, increasing register pressure. If registers spill, then memory traffic increases and code flow is disrupted. A thick loop with many temporary values may execute slower than an equivalent series of thin loops. The biggest gain will be achieved if the thick loop can be split into a series of independent loops eliminating the need to write and read temporary arrays. I found such an occasion in code 10 where I split the loop do i = 1, n do j = 1, m A24(j,i)= S24(j,i) * T24(j,i) + S25(j,i) * U25(j,i) B24(j,i)= S24(j,i) * T25(j,i) + S25(j,i) * U24(j,i) A25(j,i)= S24(j,i) * C24(j,i) + S25(j,i) * V24(j,i) B25(j,i)= S24(j,i) * U25(j,i) + S25(j,i) * V25(j,i) C24(j,i)= S26(j,i) * T26(j,i) + S27(j,i) * U26(j,i) D24(j,i)= S26(j,i) * T27(j,i) + S27(j,i) * V26(j,i) C25(j,i)= S27(j,i) * S28(j,i) + S26(j,i) * U28(j,i) D25(j,i)= S27(j,i) * T28(j,i) + S26(j,i) * V28(j,i) end do end do into two disjoint loops do i = 1, n do j = 1, m A24(j,i)= S24(j,i) * T24(j,i) + S25(j,i) * U25(j,i) B24(j,i)= S24(j,i) * T25(j,i) + S25(j,i) * U24(j,i) A25(j,i)= S24(j,i) * C24(j,i) + S25(j,i) * V24(j,i) B25(j,i)= S24(j,i) * U25(j,i) + S25(j,i) * V25(j,i) end do end do do i = 1, n do j = 1, m C24(j,i)= S26(j,i) * T26(j,i) + S27(j,i) * U26(j,i) D24(j,i)= S26(j,i) * T27(j,i) + S27(j,i) * V26(j,i) C25(j,i)= S27(j,i) * S28(j,i) + S26(j,i) * U28(j,i) D25(j,i)= S27(j,i) * T28(j,i) + S26(j,i) * V28(j,i) end do end do Conclusions Over the course of the last year, I have had the opportunity to work with over two dozen academic scientific programmers at leading research universities. Their research interests span a broad range of scientific fields. Except for two programs that relied almost exclusively on library routines (matrix multiply and fast Fourier transform), I was able to improve significantly the single processor performance of all codes. Improvements range from 2x to 15.5x with a simple average of 4.75x. Changes to the source code were at a very high level. I did not use sophisticated techniques or programming tools to discover inefficiencies or effect the changes. Only one code was parallel despite the availability of parallel systems to all developers. Clearly, we have a problem—personal scientific research codes are highly inefficient and not running parallel. The developers are unaware of simple optimization techniques to make programs run faster. They lack education in the art of code optimization and parallel programming. I do not believe we can fix the problem by publishing additional books or training manuals. To date, the developers in questions have not studied the books or manual available, and are unlikely to do so in the future. Short courses are a possible solution, but I believe they are too concentrated to be much use. The general concepts can be taught in a three or four day course, but that is not enough time for students to practice what they learn and acquire the experience to apply and extend the concepts to their codes. Practice is the key to becoming proficient at optimization. I recommend that graduate students be required to take a semester length course in optimization and parallel programming. We would never give someone access to state-of-the-art scientific equipment costing hundreds of thousands of dollars without first requiring them to demonstrate that they know how to use the equipment. Yet the criterion for time on state-of-the-art supercomputers is at most an interesting project. Requestors are never asked to demonstrate that they know how to use the system, or can use the system effectively. A semester course would teach them the required skills. Government agencies that fund academic scientific research pay for most of the computer systems supporting scientific research as well as the development of most personal scientific codes. These agencies should require graduate schools to offer a course in optimization and parallel programming as a requirement for funding. About the Author John Feo received his Ph.D. in Computer Science from The University of Texas at Austin in 1986. After graduate school, Dr. Feo worked at Lawrence Livermore National Laboratory where he was the Group Leader of the Computer Research Group and principal investigator of the Sisal Language Project. In 1997, Dr. Feo joined Tera Computer Company where he was project manager for the MTA, and oversaw the programming and evaluation of the MTA at the San Diego Supercomputer Center. In 2000, Dr. Feo joined Sun Microsystems as an HPC application specialist. He works with university research groups to optimize and parallelize scientific codes. Dr. Feo has published over two dozen research articles in the areas of parallel parallel programming, parallel programming languages, and application performance.

    Read the article

facebook twitter digg google delicious technorati stumbleupon myspace wordpress linkedin gmail igoogle windows live tumblr viadeo yahoo buzz yahoo mail yahoo bookmarks favorites email print

1