Hello James, good to see somebody interested in a benchmarking suite!
A few comments to your ideas: I'd find it very useful if you would produce one package with a (small) set of benchmarks included, prepared maybe to take more benchmarks later. I don't think benchmarks need to run as part of the tests, normally they require some customization and sometimes benchmarks simply run too long to really want to wait for them while testing the cluster. If you start with HPL this would be great. Start simple, with trying to run the HPL benchmark on only one node (but on all processors in parallel). There are a few parameters which need to be tuned such that you fill up nearly all memory of the node and get reasonable performance. To my knowledge the Goto library is the best one for HPL on x86 and x86_64. It is small and easy to build. Please also have in mind that the choice of the compiler is essential for the performance of a benchmark, so maybe you'll want to keep the build process flexible and prepared for multiple compilers (including commercial ones). Certainly, adding HPCC would be great, the HPCC results are much more meaningful than the dumb Linpack (HPL) results, they say a lot more about the performance of a machine. If you already have experience with HPCC, I hope you can add those, too. Best regards, Erich On Friday 13 April 2007 02:00, James wrote: > Hello everyone! > > I am an undergrad at Louisiana Tech University, my degree focus is on > Computer Science and I plan to graduate in November of this year. My SoC > project is the OSCAR Bench project that will incorporate benchmarks into > OSCAR. I would love to hear any suggestions, or vision you have for the > project. I expect my proposal will change a good bit between now and August. > > Some pressing thoughts I have had are: > > Where in OSCAR does OSCAR Bench belong? Should it be run as part of the > 'test cluster' if the users downloads it? Would an additional GUI button > be added if the user downloads my package? My thoughts are that the > benchmark should be able to run independently of the Wizard, but with > the requirement that the cluster be an OSCAR cluster, e.g., it has an > OSCAR database. That could be removed as well, but I would like to build > off OSCAR as much as possible, so I don't have to recreate any wheels. > This is future thinking on my part, because I am not sure what data I > may want that OSCAR has to offer (if any). > > Also what benchmarks are most relevant? If I work with the HPCC > benchmark package, do I need to include HPL? From experience I have used > the HPCC benchmark for about 4 months now, I have also built/used ATLAS > to go with it. I noticed a post here regarding IO benchmarking, so I > really want to narrow my focus down to a few specific benchmarks/packages. > > And the final thought was: Should all the benchmarks be clumped into 1 > package, or should each be its own package. For instance: an Atlas > package, a HPL package, a HPCC packages, etc..., > > Thanks again for the chance to work with you, > James Elliott ------------------------------------------------------------------------- Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT & business topics through brief surveys-and earn cash http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV _______________________________________________ Oscar-devel mailing list [EMAIL PROTECTED] https://lists.sourceforge.net/lists/listinfo/oscar-devel
