At 08:06 AM 7/15/2009, Graham, Richard L. wrote:
I need to cancel this at the last moment - my apologies. What are the items people want to address before our next face to face in two weeks ? We can do this over e-mail, or on a call.
The only thing that was hanging is the error semantics for collective operations. The question is: can an operation return the wrong result without informing the user. We've already decided that the operation may return successfully even if it has not succeeded globally. However, does this still hold if even the local result is wrong. During the last f2f Thorsten suggested that there may be protocols that can take advantage of the more relaxed semantics by allowing error and failure checking to be done outside the critical path. The results of these checks could be polled during the MPI_Validate calls. Is there a case for the utility of such optimizations? Greg Bronevetsky Post-Doctoral Researcher Lawrence Livermore National Lab (925) 424-5756 [email protected] http://greg.bronevetsky.com