[mvapich-discuss] Processor affinity (slightly special case)

[Laurence Marks]

I have some jobs (large, typically > 64 cores on a dual-quadcore
cluster) where if I run 8x8 (for instance) the master node runs into
problems because (I think, although it's hard to be 100% certain) the
mpi task is using too many resources, OS operations on it stop working
or it runs out of swap and the job never completes. One solution is to
run, for instance, 1+7+7x8 with one mpi on a single node, 7 on one and
8 on the rest. This is OK (inefficient) because I can run the single
on my head node. However, then I come to a limitation because
processor affinity on the head node means I can only run one of these
at a time, again inefficient.

I am interested in suggestions of ways around this. Some possibilities
that I can think of:
a) Is there a switch which would turn off processor affinity if only
one core is being used? (I could write a shell script to do this but
that would not be so elegant.)
b) Is there some other way to structure the job or some alternative
switches that might help.

-- 
Laurence Marks
Department of Materials Science and Engineering
MSE Rm 2036 Cook Hall
2220 N Campus Drive
Northwestern University
Evanston, IL 60208, USA
Tel: (847) 491-3996 Fax: (847) 491-7820
email: L-marks at northwestern dot edu
Web: www.numis.northwestern.edu
Chair, Commission on Electron Crystallography of IUCR
www.numis.northwestern.edu/
Electron crystallography is the branch of science that uses electron
scattering to study the structure of matter.