[Beowulf] Re: Re: Home beowulf - NIC latencies

[Rob Ross]

On Fri, 18 Feb 2005, James Cownie wrote:

[snip]

> If an ABI for MPI is so important to you and of such value to your (and
> Patrick's) clients, then there's nothing to stop you from formulating
> such a standard, or, at least starting a project to create one.
> 
> If you're right about its importance, then all the MPI implementors will
> follow your lead. If you're wrong, well, you wasted your time.
> 
> The point here is that doing this can be an informal process which
> doesn't require "The MPI Forum" (whatever that is now !?)  to endorse
> it, any more than a project on SourceForge requires endorsement by Linus
> if it runs on Linux ;-)
> 
> (Or, if you prefer, don't keep whingeing about what the MPI Forum chose
> to do, but get on and fix it for yourself).

But keep in mind that some implementations encode meaning into the values 
in mpi.h, and as a result the values aren't even the same between multiple 
instances of the same implementation on different platforms!

For example, MPICH2 encodes the size of basic datatypes in the value of 
the type.  Saves us looking around in some table for these cases (which 
are the only ones that Patrick wants us to support :)!).

So this is probably a larger problem than it seems.

Rob
---
Rob Ross, Mathematics and Computer Science Division, Argonne National Lab