--- Begin Message ---
- Subject: High-performance on AXP-Linux
- From: "Adam C. Powell, IV" <adam.powell@nist.gov>
- Date: Tue, 24 Nov 1998 12:52:15 -0500
- Message-id: <365AF24F.4BDDACF2@nist.gov>
Greetings, The Linux/Alpha platform is fast becoming acknowledge as having the best price/performance for floating point computation. This was recognized as early as around two years ago, when Digital Domain put 200 Linux/Alpha machines together to do the rendering for the movie Titanic, and more recently when the Avalon cluster at Los Alamos became the #114 supercomputer in the world at a total cost of around $300K. In the last few months, several developments have enhanced the performance of this platform, including the Free-Fast-Math library of Joachim Wesner and Kazushige Goto, and Mr. (Dr.?) Goto's hand-coded assembler BLAS (basic linear algebra subroutines) which multiply matrices faster than any other single-processor machine in the world (I think- well, at least anything else within 10 times the price). Johannes Hausmann has put these and other fast linear algebra routines into a single fastmath/BLAS/LAPACK RPM builder which greatly simplifies installation of these disparate pieces. Richard Payne has put together a page to summarize these and other high-performance efforts at http://www.alphalinux.org/docs/high_perf.html. In response to that page, it was suggested that a mailing list be created to discuss these topics. This list has been created, and is open to new members. To join, simply send me email. (If you reply to this, make sure *my* email address is in the To: field, not a list address.) Possible discussion topics might include: * Tuning of existing routines for improved speed, e.g. block size issues, performance on 21164PC sans on-chip L2 cache, etc. * New routines to accelerate FFTs, sparse matrix solvers, FEM/BEM matrix construction, etc. * Requests for help with coding of inner loops of custom codes (neural nets? spreadsheets? mozilla? quake? :-), including assembler instruction scheduling, cache issues, etc. I suspect the list will focus on scientific computation, but any high-performance topic is welcome. Note: this list is based on decade-old non-majordmo list technology, and list updates happen once a day, and if I am unreachable, an update request could potentially take several days to be processed. It will not have an archive right away (though a volunteer has stepped forward to create one soon), but I will save all messages and will gladly honor forwarding requests. If anyone has a better list server available, then perhaps the list should be hosted there; in the meantime, I'm making available what I have. Cheers and happy hacking! [P.S. If anyone's on debian-alpha, feel free to forward this.] -Adam `Cold Fusion' Powell, IV http://www.ctcms.nist.gov/~powell/ ____ USDoC, National Institute of Standards & Technology (NIST) |\ ||< | Center for Theoretical and Computational Materials Science | \||_> |
--- End Message ---