[Beowulf] MPICH vs. OpenMPI

[stephen mulcahy]

Prentice Bisbal wrote:
 > Can anyone give me a quick comparison of OpenMPI vs. MPICH? I've always
 > used MPICH (I never liked lam), and just recently heard about OpenMPI.
 > Anyone here using it?
 >

When we initially built our cluster for ROMS modelling in 2006 we were 
using LAM-MPI but ran into some problems with it (there was a known bug 
in relation to NFS which we hit) we moved to MPICH2. It was stable and 
worked pretty well.

Once we had the cluster bedded in and things were running smoothly, we 
took another look at the performance of the various MPI libraries and 
came up with the following numbers for ROMS running on our cluster -- as 
usual YMMV, and I haven't tried to reproduce this with the latest 
versions of these libraries so the results may be entirely different now 
(is there a standard list disclaimer for these kind of discussions? :)

The numbers are the time it took our model running under ROMS to 
generate a single history file (checkpoint), lower is better.

MPICH2 1.0.5p3 34 min
LAM-MPI 7.1.3  41 min
OpenMPI 1.2    31 min

So for us, it OpenMPI was clearly the right library to go with.

Other nice things about OpenMPI include the fact that it seems to have 
some built-in support for affinity and from a user perspective it is 
pretty user-friendly (you don't need to start up separate daemons before 
starting your job and it seems to do a pretty good job of cleaning up 
after itself).

hth,

-stephen

-- 
Stephen Mulcahy, Applepie Solutions Ltd., Innovation in Business Center,
GMIT, Dublin Rd, Galway, Ireland.  +353.91.751262  http://www.aplpi.com
Registered in Ireland, no. 289353 (5 Woodlands Avenue, Renmore, Galway)