I find it ironic to see a university researcher cautioning against the
"researchiness" of proposals that are pushed by vendors or national lab
people. MPI2 is by no means an academic or research exercise (at least not for
vendors). With the possible exception of one chapter, all proposals are
driven either by the desire of some vendors to take advantage of their
hardware or strong pressure from the user community to agree on standards.
MPI-IO is driven by strong user community pressure for a standard on parallel
IO, and the existence of parallel IO libraries. Dynamic is driven by pressure
of the research community (although I personally believe it to be the least
useful extension to MPI, since we have no good concept of how to mix the "user
control resource management" view of PVM and the "system control resource
management" view of any vendor system). One-sided is driven by the recognition
that memory coupling, without full shared memory support, is increasingly
widespread (Cray, Dec clusters, SGI clusters, etc.). Misc is driven by
requirements of library developers. And so on.
On the other hand, I find it very pleasing that Greg states that MPI-1 was
very much standardizing current practice. As a person that participated in
MPI-1 I can testify that this was far from from true. When we started, there
was nothing ressembling MPI communicators, nothing ressembling MPI datatypes,
nothing ressembling attribute caching, nothing as extensive as MPI
collectives, and so on. I think it is a good testimony to the success of MPI
that people believe that MPI1 standardized existing practice.
That said, 300 new pages of MPI text should worry us (even if the final text
is likely to be much shorter, when discussions, alternatives, repetitions,
etc. are pruned out). We should carefully think how we want to package MPI2,
what is part of MPI, and what is part of separated, stand-alone libraries
(MPI-IO?).
Marc Snir
IBM T.J. Watson Research Center
P.O. Box 218, Yorktown Heights, NY 10598
email: snir@watson.ibm.com
phone: 914-945-3204
fax: 914-945-4425
URL: http://www.research.ibm.com/people/s/snir