From: Bjorn Lindgren <bjorn@no.sp.am>
Newsgroups: comp.parallel.mpi,comp.parallel.pvm
Subject: Re: Dual- vs. single processor motherboards
Date: 20 Mar 1999 22:55:11 GMT
Organization: UNIX: Live free or die
Sender: Bjorn Lindgren <bjorn@nexus.swip.net>
Message-Id: <7d190f$1pa$1@news.kth.se>
References: <UTkG2.4447$_n2.99958@carnaval.risq.qc.ca>
    <36EAA3B0.47FC0EB9@gci.net>
Mime-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Transfer-Encoding: 8bit
User-Agent: tin/pre-1.4-981225 ("Volcane") (UNIX) (Linux/2.2.3 (i686))
Xref: ukc comp.parallel.mpi:4789 comp.parallel.pvm:8173


In comp.parallel.pvm Curt Timmerman <ctimmer@gci.net> wrote:
> With a single process I could get 99%+ utilization on the CPU. Running 2 process
> on the same node, both CPU's would also reach 99%+ utilization. The tests I ran
> probably allowed the processors to run mostly from L2 cache. As data
> requirements increases, main memory access would lower the utilization percents.
> You don't have to write multi-processor application to get most of the benefit
> of the dual processor setup.

> I too found that the message passing overhead was the real bottleneck. In an
> attempt to speed things up I purchased a second ethernet card for each node. The
> second network is devoted strictly to PVM message passing. Using this with a
> switching hub should improve things. I haven't run any tests yet.

Using POSIX Threads you only need to spawn one thread on each node to take
advantage of both CPU's, and lowering the network usage with 50%.

- Bjorn Lindgren