[Beowulf] Alternative to MPI ABI

Erik Paulson epaulson at cs.wisc.edu
Tue Mar 22 11:17:40 PST 2005


On Tue, Mar 22, 2005 at 07:22:36PM +0100, Joachim Worringen wrote:
> 
> Largely, what you describe sounds more like some variant of grid 
> computing. MPI is not about grid computing, it's mostly about 
> performance, but can be embedded in a grid environment. But it should 
> not try to represent a grid environment itself.
> 

I think the point Don is trying to make is that MPI was designed to 
be a parallel environment for machines that could be tightly-coupled
MPPs or piles of machines in racks or shelves - and anything that 
couldn't be done well on both sides was left out of the standard. 

The situation on the ground today is that effectively all that is left
is piles of machines - some with fast interconnects, but most with just
ethernet, and all with a higher-than-we'd-like failure rate. But we're
still constrained with a programming model that is designed to take the
same program and run it efficiently on a T3E or a cluster of PCs. I'd
also bet that if we looked at most MPI programs out there, they 
really could be restructured very easily to take advantage of a dynamic
environment, if MPI properly exposed that to them (and MPI-2 is _not_
good enough) 

>From looking at Don's earlier post, I was thinking that PVM gave us a lot
of what was listed (and the LAM underneath LAM/MPI, and HARNESS, etc).
If the OpenMPI effort decides to revise some of the MPI standard,
I hope they focus on exposing the cluster reality that most parallel
programs run on, and don't worry so much about making it run efficiently
on machines that no longer exist.

-Erik



More information about the Beowulf mailing list