[Mpi-forum] Missing MPI primitives - Probe+Wait

N.M. Maclaren nmm1 at cam.ac.uk
Tue Apr 20 08:29:05 CDT 2010

On Apr 20 2010, William Gropp wrote:

>I just wanted to respond to this.  In fact, the Forum was well aware  
>of how incomplete the POSIX thread standard was and was careful to not  
>mandate them (!) - we did receive suggestions from people who thought  
>MPI codes would be *more* portable if we said MPI must support pthreads.

I certainly didn't mean to imply that it wasn't - the specification makes
it very clear that it was!  All an implementation is required is to tell
the programmer what it supports, which suits me - the answer might even
be correct!

Similarly, the issues of progress, scheduling and memory consistency were
clearly understood, as the remarks on MPI-2 one-sided communication make
clear.  It also makes it clear that hairy things like IEEE exception flags
and signals are "here be dragons".

One of my main reasons to stick to MPI-1 is that it is as near to
bulletproof as one can hope to get nowadays.  I have come across very,
very few MPI implementations where facilities deliver the syntax but
not the semantics, and then only in ways that were simple bugs, and not
design restrictions.

Nick Maclaren.

More information about the mpi-forum mailing list