Newsgroups: comp.parallel.pvm
From: ce107@hydra.cfm.brown.edu (C. Evangelinos)
Subject: Re: workstation clusters
Organization: Brown University
Date: 04 Sep 1996 05:35:03 GMT
Message-ID: <CE107.96Sep4013503@hydra.cfm.brown.edu>

In article <50hkq3$7vg@netnews.upenn.edu> jeff@adeno.wistar.upenn.edu
(Jeff Taylor) writes: 

>This card looks great from the specs but the price is higher than the
>PC I was going to plug it into!  For the $2000 you'll have to spend
>on this card you could get another complete node, which even if you
>have to use the slow 100Megabit ethernet connection will probably
>give you better performance (more nodes = faster).  
			      ^^^^^^^^^^^^^^^^^^^
This is only sometimes true (and even on tightly coupled MPPs one can
find examples where just the opposite is true). It depends on your
algorithm and your partitioning stategy, the problem size per
processor and of course the added amount of communication you may have
to do as compared to the gain in computation speed. The only certain
advantage of the extra node is in terms of throughput of serial jobs.
One more node means one more free workstation. :-)

>It's too bad
>really, I would like to build a system based on a switch such as the
>Parastation but for now it's out of my price range.

Well, the good thing seems to be that fast interconnects for clusters
of workstations (or SMP boxes) are in fashion lately: Memory Channel
from Digital, SCI Cluster from Dolphin just a few days ago and this
Parastation today - if the situation in this newsgroup reflects the
trends out there we should expect better (and cheaper) hardware and
software pretty soon, supporting some form of DSM as well.

Constantinos Evangelinos
Center for Fluid Mechanics
Brown University/Division of Applied Mathematics

PS> Just to avoid any misunderstandings I'm completely unrelated to
the Parastation project/company etc. in fact I heard about them just
today from this newsgroup.


