HZ and HPC

Ken Chase math at velocet.ca
Wed Jan 29 06:10:16 PST 2003

On Wed, Jan 29, 2003 at 08:55:11AM -0500, Ken Chase's contradicting
himself again with...

> With large caches today you can tune a class of jobs for your cluster
> that are below the threshhold size nicely, and run those jobs differently
> from the others (using different #s of nodes, possibly even different
> interconnects as thrashing will obviously destroy scaling).

Actually, now that I think of this, it may well do the opposite in fact 
if you have a job that runs slower for some reason, the relative speed
of your network is now faster for that job - it may well allow you to
maintain scaling with slower interconnects between nodes. Most likely
you may be able to scale the same amount over a few more nodes than
with the job running faster on the CPU.

(Consider the pathological condition - if a job ran 100 times faster on
a CPU when not thrashing the cache, you'd need a much faster interconnect
(much lower latency) to get the messages to the other nodes in time to
avoid having them waiting on data.)

Sometimes I think that really wide pipes that are otherwise cheap and slow
might serve heavily loaded (multiple jobs) CPUs more efficiently. Comments?

Ken Chase, math at velocet.ca  *  Velocet Communications Inc.  *  Toronto, CANADA 

More information about the Beowulf mailing list