[Beowulf] AMD64 results...

Michael Will mwill at penguincomputing.com
Thu Dec 16 10:16:35 PST 2004


On Thursday 16 December 2004 08:12 am, Craig Tierney wrote:

> > don't forget to turn on node interleave in the bios, too.
> 
> Why?  If you are planning to have a single process access
> the memory of all of the nodes (cpus) then yes.  If you are
> running MPI jobs or multiple processes that stay local to their
> own memory, they don't you want bank interleave on but node
> interleave off?
> 
> I have seen better performance for MPI jobs, 1 process per
> cpu, with node interleave off.

This assumes you have a 2.6 or numa kernel that makes sure that
the process stays on the CPU with the memory. A year ago I did 
benchmarking on SLES8 on a dual opteron and when running it
over and over again I would observe two very distinct results,
depending on if the process and its ram were on the same CPU or not.

Turning on node-interleave averages the results out, which is not
needed in real life except for when you need to have consistent publishable
results, or are worried about timing variations making your debugging harder.

Michael

> Craig
> 
> 
> > 
> > > Kudos for the pathscale-1.4 compiler with -O3.
> > 
> > ironically, icc -xW generates pretty good-for-opteron code,
> > though of course, it's 32b.  I haven't tried using icc to
> > generate em64t/and64 code.
> > 
> > regards, mark hahn.
> > 
> > _______________________________________________
> > Beowulf mailing list, Beowulf at beowulf.org
> > To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
> 

-- 
Michael Will, Linux Sales Engineer
NEWS: We have moved to a larger iceberg :-)
NEWS: 300 California St., San Francisco, CA.
Tel:  415-954-2822  Toll Free:  888-PENGUIN
Fax:  415-954-2899 
www.penguincomputing.com




More information about the Beowulf mailing list