I sent couple e-mails in hopes to discuss some points in the
MPI-2 document concerning threads and MPI that seem to be
inconsistent and/or misleading.
Seems like not many people are interested in making sure that
interaction of POSIX-compliant thread libraries and MPI-2 is properly
defined. This is quite surprising since the importance of thread
usage in MPI implementations as well as MPI applications
seems to be quite obvious. Indeed, I do not aware of a portable
alternative to threads for implementation of the "asynchronous agents"
that are necessary for one-sided communication, message
handlers, and other things in MPI-2. Needles to say that efficient use
of multi-processor hardware by MPI applications, as well as implementation,
can be easily done with threads. Note, that MPI implementations themselves
can use multiple processors to run faster; this is impossible in a
single-threaded MPI.
I do not think it is necessary to resend the e-mails; if it is, I got the
"improved" versions.
So, I still hope that my remarks will be considered, since the current
version of chapter 7 and 13 are not in the perfect shape.
I would like at least to hear from the person who is developing these
chapters for the Forum.
Thanks, Boris Protopopov.