As the page URI seems to be
diffusing rapidly through the web, it becomes INCREASINGLY IMPORTANT for
me to warn visitors: This project is under construction! I've
literally changed every page linked below two or three times today. I
may end up rerunning some numbers tomorrow or Friday. Do not
assume that everything is correct and that these numbers are final.
Look them over and enjoy, but come back next week to see the final
numbers (which may still be incorrect).
This page contains the running results of a series of benchmarks being run on a dual athlon the week of 3/12-3/16 (2001). I'm posting things as I get them, so people have time to comment on any results that are insane (as I am not personally familiar with "sanity" on all the benchmarks that I've been asked to run). If the comments are communicated in a timely fashion I'll try to rerun anything that isn't working correctly and post the corrected results.
Some of the numbers posted for comparison purposes aren't really very good for that purpose. For example, the lmbench page compares a 933 MHz PIII running the 2.2.16 kernel to the 1.2 GHz dual Athlon running the 2.4.2 kernel. It is very difficult to connect e.g context switch rates or certain latencies to the system hardware (CPU and memory subsystem) versus the system software (kernel and libraries) where they are clearly different. Remember, I don't own the Athlon I'm testing and have to run what is there, so such mismatches are inevitable.
Disclaimer: I'm not being paid by AMD to run these tests. In fact, I'm not being paid by anyone to run these tests. Of course, if anyone wants to pay me to run these tests, I'd be happy to send them a P.O. box number (small unmarked bills in a brown manila envelope would be most convenient). If cash is difficult just now, feel free to send hardware or tee shirts.
An Important Warning! I do not certify that these results are correct or suitable for use as an accurate representation of comparative systems performance. The results may be useful to you (obviously I hope that they will be) but are not guaranteed to be suitable for any purpose whatsoever. The system itself is a beta system with a beta 760MP chipset. It is running 2.4.2 which is a fairly new kernel. I have no control whatsoever over kernel, compiler or hardware configuration. Use these results or don't use them at your own risk to make hardware purchase or design decisions.
Free (and well-earned) Plug! The machine these tests were run on was provided by ASL Computers, which is a company that sells pre-installed linux configurations (only, as far as I can tell - WinXX users need not apply) ranging from desktops to full racks (both 2U and 1U form factors). They also sell a whole range of peripheral stuff, RAID enclosures and so forth. The company looks to be located in the bay area of California. If you're looking for quotes on beowulf (or generic cluster or even personal desktop) hardware it can't hurt to look over their offerings and/or contact them.
The following are links to the sources used to test the dual athlons. I am hoping that these evolve into something of a "beowulf performance suite". Note that in some cases it is not possible to fully "package" the sources, as they have dependencies not only on other tools but on particular builds of other tools. Several of them, for example, require LAM MPI built using the --with-rpi=usysv option. I do my best in these cases.
In all the cases where the tool used has its own primary source and maintainer, both the original/primary source and the packaged or modified source used in the benchmark are included below. On the benchmark page some effort will be made to detail the changes made (if any) so that readers can judge for themselves (and tell me!) whether the results are still properly comparable to those produced by the original sources.
I am publishing these sources, collected in one central location for easy retrieval, at least in part as a "request for comments" page. So please feel free to comment. If you just find it useful to have such a single site collection, let me know that as well.
These sources should (in many cases) form a decent collection of single cpu system benchmarks. Beowulf engineering can critically depend on a knowledge of certain microscopic bottlenecks and rates, but it isn't easy to determine the best set of tools to perform those measurements and harder still to collect them and build them and apply them with some assurance that your numbers are meaningful and comparable to other such numbers evaluated on other platforms.
Please feel free to retrieve and apply these tests in your own environments and please report to me any places where you think that they are broken or producing erroneous results.
Note that some of these links are still inactive as the sources either cannot be packaged or have not been packaged. Eventually I'll probably get through them, but I do have a job and work to do on top of benchmarking and beowulfery.