Crisis in HPC Discussion - Nick Maclaren, University of Cambridge

Newsgroups: uk.org.epsrc.hpc.discussion
From: nmm1@cus.cam.ac.uk (Nick Maclaren)
Subject: Re: Crisis-in-HPC conclusions
Organization: University of Cambridge, England
Date: 4 Oct 1995 13:25:15 GMT
Message-ID: <44u1vr$amn@lyra.csx.cam.ac.uk>

In article 37, lyndon@epcc.ed.ac.uk (L J Clarke) writes:
|> IMHO this posting is spot-on. Can we disentangle the performance and
|> software issues in using the parallel machines from the performance
|> and software issues in using the newer faster commodity microprocessors
|> (superscalar, superpipeline, cache, etc)? I do hope so, in fact I believe we
|> must, and from time to time we may not have. These microprocessors appear
|> in parallel machines (esp. MPP type) because these microprocessors appear
|> in workstations.

Well, we could, but would it be a good idea? This mistake was made back in the 1970s, when supercomputers started to go vector (usually driven by library calls from Fortran) and commodity computers started to exist (and were driven by assembler, leading to C). We are still suffering from the legacy of this split.

As far as the caching problems go, these are EXACTLY the same for a single microprocessor and a parallel system, with different scaling parameters. The algorithmic constraints are almost identical.

Nick Maclaren,
University of Cambridge Computer Laboratory,
New Museums Site, Pembroke Street, Cambridge CB2 3QG, England.
Email: nmm1@cam.ac.uk
Tel.: +44 1223 334761 Fax: +44 1223 334679