I propose that we revise the channel specification in the RT chapter. During
the RT subcommittee meeting last week, we came very close to eliminating the
non-bufpool variants of the channel creation and modification routines. I
propose that we go ahead and delete all of the standard buffer versions and
rename the bufpool versions by removing the word "BUFPOOL" from the name.
(E.g., the current MPIRT_CHANNELS_INIT() would be deleted, and
MPIRT_CHANNELS_BUFPOOL_INIT() would be renamed to MPIRT_CHANNELS_INIT().)
Furthermore, I recommend that we state in the description of the channels that
MPI/RT will provide no other buffering for a channel other than that explicity
provided by the user via the bufpool. (For example, if all of the receive
buffers have been exhausted, any further requests at the send side will not
complete until one of the receive buffers is consumed.) I think that only by
eliminating extra buffering can we hope to meet real-time contraints in general
and the needs of resource contrained systems in particular. Unless there are
objections, I would hope to see these changes incorporated into the version of
the RT chapter which appears at SC96.
--Jerrell
--
Jerrell R. Watts
Scalable Concurrent Programming Laboratory
California Institute of Technology
jwatts@scp.caltech.edu -- http://www.scp.caltech.edu/~jwatts