p4 v itinium

Donald Becker becker at scyld.com
Fri May 17 14:23:00 PDT 2002

On Fri, 17 May 2002, Josip Loncaric wrote:

> ...
> based on domain decomposition, the optimal amount of memory per node
> grows very quickly as a function of the CPU/communication speed ratio.

This comes as a surprise to many: as processor performance increases:
  the per node communication requirement grows less rapidly than expected, but
  the per node memory requirement grows faster than expected.

> If you double the CPU speed but keep the same network, you can maintain
> performance if your domain volume/surface ratio doubles as well.  This
> can happen if the volume (computations) grows by a factor of eight and
> surface (communications) by a factor of four.  In other words, per node
> memory requirements can grow with the cube of the CPU/communication
> speed ratio.

This paragraph should be the canonical response when people ask about
the future communication and memory needs on a cluster.  Not every
problem is a physical finite element grid, but that's the easiest mental
model when explaining trends.

Donald Becker				becker at scyld.com
Scyld Computing Corporation		http://www.scyld.com
410 Severn Ave. Suite 210		Second Generation Beowulf Clusters
Annapolis MD 21403			410-990-9993

More information about the Beowulf mailing list