Re: Is non-blocking collective I/O desirable?
Linda Stanberry (email@example.com)
Tue, 18 Feb 1997 08:08:31 -0800 (PST)
Thanks to Nick for verbalizing the puzzlement I have had both about
nonblocking collective I/O and the proposed generalized request
mechanism for allowing the creation of new nonblocking I/O operators.
It seems to me that MPI assumes some mechanism for implementing more
than one logical thread of execution (not necessarily a POSIX thread,
however) for MPI-1. As Nick points out, this can be implemented using
third party or interrupts, if threads are not available or desirable.
MPI-2 merely extends this model, and both collective nonblocking I/O
and GRs are logical (to me) extensions that are consistent with the
rest of the MPI-2 functionality.
The real issue appears to be getting the synchronization semantics
correct for these extensions. But doesn't MPI-1 put that burden
squarely on the user, which is reasonable since the user knows the
dependencies of operations and MPI implementations don't? That is,
it's the user's challenge to get the synchronization right, and MPI's
challenge to provide the tools that will enable users to achieve this.