* Anthony Skjellum <[email protected]> [2013-08-26 14:29:01]:

If I understand this correctly, some of us suggested allocate-on-receive, free-on-send mode, but left this out in MPI-1. I advocated this type of reactive semantics ... That's how the Reactive Kernel worked at Caltech for instance... Allowing for MPI to choose efficient memory for messages of some sizes can provide nice properties, provided the memory is good to compute in. There are always concerns about starvation from holding the messages too long :-)

I'd like to see this get full consideration. It could be a big help in exascale to have fine-grain computing.

For instance, could we not get zero-copy unexpected short messages in this way?

Don't forget to consider one-sided communication. Many of the proposed uses for MPI_Arecv could be implemented using one-sided, given some more application knowledge (which the application programmer has). For example, implementing zero-copy unexpected short messages is pretty trivial using the newer one-sided functions. Dries