I have mentioned that I think the proposal for removing the send buffer access restriction is worth examining. I have even said I think it likely (not obvious) that the rule should be relaxed. First we should conclude that keeping it has trivial potential for being important to some MPI implementation in comparison the burden it puts on application writers. Performance matters a lot and always will.
I thank those who provided examples of situations where relaxed send buffer rules would be helpful.
>From the start of this discussion I have assumed there would be defensible parallel algorithms that tempt the programmer to violate the rule. I have also assumed there are probably some programmers who violate the rule, either because they do not know about it or because they think they know enough to be sure the rule is irrelevant. I did not have a good estimate of how often the rule is a potential problem but I did see how it could matter.
Finally, I think that unless the problems the send buffer rule creates for MPI application writers are real and significant, we should lean toward keeping it. If we let it go to solve hypothetical problems and later develop an architecture that could really exploit the freedom the rule provides to implementors, we will be unable to get the rule back.
Dick
BTW - I do not see that the decision about the send buffer rule has much connection to "Proposal EH2: add const keyword to the C bindings". In my opinion, the two discussions should be kept apart . I will tackle the "const" issue in a new thread. The send buffer rule seems to have become the dominant topic of this thread.
Dick Treumann - MPI Team/TCEM
IBM Systems & Technology Group
Dept 0lva / MS P963 -- 2455 South Road -- Poughkeepsie, NY 12601
Tele (845) 433-7846 Fax (845) 433-8363