Path: utzoo!utgpu!jarvis.csri.toronto.edu!cs.utexas.edu!samsung!usc!ucsd!ucsdhub!hp-sdd!ncr-sd!ncrcae!hubcap!Jim
From: jim@pythagoras.doc.ic.ac.uk (Jim Crammond)
Newsgroups: comp.parallel
Subject: Re: timing parallel programs
Message-ID: <8235@hubcap.clemson.edu>
Date: 5 Mar 90 13:52:13 GMT
Sender: fpst@hubcap.clemson.edu
Lines: 13
Approved: parallel@hubcap.clemson.edu

In my experience you can get pretty consistant timings on a normal system
with having to use tmp_affinity. This is, I think, because a running processor
will not be interrupted if there is an idle processor available to respond
to interrupts and other distractions. So all you need is a quiet system.

I have used the "cluster" facility on the BBN Butterfly - I find it a pain.
It does not take many users to take away most of the machine and they
always take more than they need; what's particularly annoying is when people
run benchmarks in a large cluster and logout - many a time i have seen that
the benchmark has apparently finished but the cluster isn't released.
In my opinion a case of a good idea in theory but not in practice.

-Jim.