Re: IBM 390
On Thu, Mar 25, 1999 at 10:34:37PM -0500, Phillip R. Jaenke wrote:
>
> Michael; I've got a standing bet with my boss that if I get an SP, I can
> have it outperforming a Beowulf cluster with identical x86 processors and
> an identical number of nodes, at less than 3/4 load. It's not a matter of
> brute speed; it's a matter of architecture, as I see it.
>
Very true for many problems. Don't knock Beowulf clusters though. Let the
press overstate the abilities of a non-commercial solution for a change.
For problems that can be easily partitioned into many pieces, without
requiring a lot (very relative term) of communication Beowulf clusters
work really well. For problems that require a lot communication (that
can even mean needing to share a huge block of memory) then Beowulf
clusters just don't cut it.
Jay Treacy
Reply to: