Re: BLAS recommendations SUMMARY
On 22 Dec 1999, Camm Maguire wrote:
> Emil Briggs <briggs@tick.physics.ncsu.edu> writes:
(And a whole lot of other people wrote)
Thanks! That looked like a very, very useful summary. Although the
thread will now of course be in the list archives from which it can
easily be retrieved (via the www.beowulf.org page) I took the liberty of
grabbing the original question and all the responses and putting them
"as is" on a text page (links not active, sorry, as I didn't want to
mess with html reformatting of the responses). It can be retrieved at:
http://www.phy.duke.edu/brahma/blas
for anyone who wants a text copy of the whole discussion and all the
extremely useful suggestions. If any more come in, I'll try to remember
to append them.
To summarize my own interpretation of the responses:
http://www.netlib.org
has lots of LAPACK and BLAS stuff, including sources, for levels 1, 2
and 3 which basically describe vector, vector/matrix, and matrix/matrix
transformations.
http://www.netlib.org/blas/faq.html
is a BLAS FAQ, written (apparently) by Greg Henry, who also provides a
non-GPL but still basically free binary link library optimized for the
PPro and PII, but not fully optimized for the PIII or Athlon (no
prefetch).
The ATLAS package:
http://www.netlib.org/atlas/
is (according to Emil Briggs of NCSU, <briggs@tick.physics.ncsu.edu>,
who seems to be something of an expert on this) is "a good solution
under Linux". The license isn't quite GPL, but shouldn't obstruct its
usage.
Emil also has optimized at least level 1 things for the Athlon. The
result might be adaptable to the PIII prefetch fairly easily. The
result is at:
http://nemo.physics.ncsu.edu/~briggs/blas_src_v0.11_tar.gz
Camm Maguire <camm@enhanced.com> pointed out that Debian has an
atlas-based package that provides a binary compatibility blas library.
He also backed his assertion that atlas is a very good solution with a
set of benchmarks on a cluster that showed what appeared to be very
reasonable performance scaling (although I'm still trying to figure out
just what the benchmarks mean...;-). This would support the assertion
that ATLAS works well under linux and can be packaged for distribution
without running into licensing issues.
Finally, I myself discovered (as was also pointed out by several
respondants) that LAPACK/BLAS routines are available via the RPM
repository:
http://rufus.w3.org/linux/RPM/
In particular, BLAS man pages are available there, which should prove to
be invaluable even to those who e.g. use Greg Henry's optimized
libraries.
To close with a mildly political statement, I'd recommend that in cases
where folks have energy to put into LAPACK/BLAS for linux and
beowulfery, they should focus on open source solutions (GPL or not) as
ATLAS appears to be and avoid the "easy" route of using binary libraries
with uncertain licensing. If just a few folks who are interested join
Emil in his efforts to optimize for the common processor families it is
likely that the libraries will very rapidly become extremely good, while
in the case of closed/proprietary libraries (however nice they may be)
you are always at the mercy of the owners with respect to bugs and
performance both. Remember, a beowulf is a COTS, OPEN SOURCE based
cluster supercomputer, at least to the extent possible.
rgb
Robert G. Brown http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567 Fax: 919-660-2525 email:rgb@phy.duke.edu
Reply to: