[MITgcm-support] MITgcm on cluters!

[Dimitris Menemenlis]

Van Thinh, one problem that I have seen in the past with clusters is poor 
implementation or tuning of MPI.  If you have an expert nearby, you can ask them 
whether they can tune MPI to make sure that there is no MPI message passing 
contentions between processors that live on the same node.  You can also try 
running using a single CPU in each node.  That will eliminate MPI contention 
between processors on the same node, if that is the cause of your slowdown. 
Another way to reduce this problem is to run in mixed memory mode, shared memory 
(two threads) on each node and MPI across nodes.  Dimitris

-- 
Dimitris Menemenlis <menemenlis at jpl.nasa.gov>
Jet Propulsion Lab, California Institute of Technology
MS 300-323, 4800 Oak Grove Dr, Pasadena CA 91109-8099
tel: 818-354-1656;  fax: 818-393-6720