Re: MPI and Threads

Rusty Lusk (lusk@mcs.anl.gov)
Mon, 17 Mar 1997 15:08:40 -0600

| I gotta disagree. But to some extent this is a moot issue. I see no benefit
| at all to officially blessing any one threads package in the MPI
| standard. If users want pthreads in sufficient numbers then the market will
| adapt to fill this need. And if they don't, then we would be better off not
| attempting to dictate otherwise in the standard.

I gotta agree. One reason that we should not require or endorse any
particular thread syntax, and instead focus on the semantics as we have
been doing, is that for MPI, the semantics is the point. To make this
explicit, one can envisage a system that has a POSIX-compliant pthread package
that does *NOT* follow the rules in our "MPI and Threads" discussion, but also
has another thread package that does do so. Then the programmer must not
use pthreads, but the threads that work well with MPI. The program might
not be as portable as it was before the programmer introduced explicit
multithreading, but that is the fault of the threads packages, not MPI.

Rusty