[Beowulf] Really efficient MPIs??

[stephen mulcahy]

Jeffrey B. Layton wrote:
> amjad ali,
> 
> If you are going to use GigE (TCP), I would recommend Scali MPI
> (www.scali.com). It's commercial, but it's the best MPI I've ever tested
> (the fastest). Plus it works with TCP, Myrinet, and IB without having
> to recompile.
> 
> If you don't want to pay money for an MPI, then go with Open-MPI.
> It too can run on various networks without recompiling. Plus it's
> open-source.
> 
> Both MPI's use processor affinity, so that will help improve performance.
> I think Scali MPI is better - it has great profiling tools, tuning 
> capability,
> and it's really fast! But it don't want a commercial MPI, then go with
> Open-MPI. I know the head developer of Open-MPI and I know most
> of the Scali MPI team as well :)

For our model we did some initial performance testing with MPICH2, 
LAM-MPI and OpenMPI and found OpenMPI to have a performance advantage of 
a few percent over the others for our situation. So I'd second the 
suggestion of giving OpenMPI a look.

We also considered testing ScaliMPI but at the time it didn't support 
Debian based Linux systems (which ours is) so it wasn't suitable (I'm 
aware of alien and friends, the effort required to bring Scali to the 
point of being testable on our system wasn't deemed worth it at the time).

-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)