It seems to me that a fundamental difference between this proposal and
the current one in the chapter is where the req_state (called
extra_state in the chapter) is given. Otherwise, this new proposal
does not seem much different than the old persistant version (which
was one of the previous options). The placement of the req_state has
been discussed by the group. Based on the e-mail discussions where
everyone who participated agreed that the req_state should be setable
each time a GR starts, it was placed in the start (MPI_GR_BEGIN) call.
Not to rehash all the reasons, but generally it was felt that many
uses would involve setting up a GR template (to do an I/O operation,
for example) and then it would be started many times for different
parameters (such as file). Because of this desire, you cannot use the
current MPI_Start function because it has no parameter to pass
req_state. This is one of the primary reasons that the current
proposal is written the way it is. I would like to know if there are
others who support moving the req_state. If not, then I think the new
proposal needs to be rethought since one of the stated advantages of
simplicity and similarity to current MPI-1 may be diluted. If there
is support for this change, then I would like to continue the
discussion on other topics where the exact sematics of the new
proposal in certain situations is not completely clear to me (maybe it
is like the current one so was unstated).
Steve