[mvapich-discuss] MPI3

Jens Glaser jglaser at umn.edu
Wed Oct 17 20:51:59 EDT 2012


Hi,

I am curious whether the MPI-3 extensions to one-sided communication are presently or will be implemented in MVAPICH2.

In particular, I think I have a need for dynamic windows (MPI_Win_create_dynamic, Win_attach, Win_detach),  since I want to copy multiple send buffers to the same buffer on the receiving
side, and the offsets are only known on the sending side. However, the buffers I am planning to use as the RMA window
are not static during the runtime of the application, they can be reallocated. The static MPI-2 solution for this, MPI_Win_create requires collective communication on the whole communicator,
and it would be slow with a large number of processors if it is performed every communication step. On the other hand, the group of processors effectively required to sync in a communication step is only a subset of the whole communicator, namely the cell neighbors of a domain-decomposition scheme. So if MPI_win_attach/detach would be lightweight operations to attach/detach the receive buffer, and if they could be performed in every communication step, I think the proposed MPI3 extensions would very likely solve my problem.

I saw the paper "Design and implementation of key proposed MPI-3 one-sided communication semantics on infiniband" regarding MVAPICH2 and I am wondering
whether the implementation described therein is already available, or at least, available for testing purposes.

Thanks,
Jens.




More information about the mvapich-discuss mailing list