[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Machine clusters & Axiom (was: Re: [Axiom-developer] Re: [Gcl-devel]
Re: Machine clusters & Axiom (was: Re: [Axiom-developer] Re: [Gcl-devel] gcl sockets)
18 May 2004 16:07:01 -0400
Gnus/5.09 (Gnus v5.9.0) Emacs/21.2
David MENTRE <address@hidden> writes:
> Hello Tim,
> root <address@hidden> writes:
> > (b) in the long term there is a sub-effort to put Axiom on a beowulf.
> > I have access to a 128 node cluster with 2Ghz processors on each node.
> > Of course getting Axiom running in small-grain parallel mode (as
> > opposed to an independent Axiom on each node) is a hard problem
> > on its own. I've done some initial experiments with a matrix domain
> > (parallel matrix multiply) but not much since then. I was hoping to
> > have a student to work on this but that fell thru.
> As a side note, I've done my PhD in a research team looking at those
> cluster issues (and my own PhD was on the shared memory topic). I doubt
> that you can gain any speedup on a beowulf-like cluster for fine grain
> computations, due to high cost of communications (to much latency, to
> little bandwidth). However, coarse grain parallelism (e.g. one long
> Axiom computation par node) seems to me much more doable.
Agreed. In general, I know of no broad area of symbolic computation
which would present an easy basis for coarse grain factorization. And
the systems we are discussing don't seem to do numerical work much,
though if they did, the obvious thing to do would be to interface with
blas/lapack interfaces, which can be made binary compatible with
automatically tuned versions (atlas, which can be multithreaded) and
possibly parallel versions, e.g. pblas, scalapack.
> David MENTRE <address@hidden> -- http://www.nongnu.org/axiom/
> Axiom-developer mailing list
Camm Maguire address@hidden
"The earth is but one country, and mankind its citizens." -- Baha'u'llah