Hi all, The Hardware Topologies Working Group wants to announce the first vote for the MPI_Cart_create_weighted / Topology aware Cartesian communicators:
Annotated pdf as read in the June 2020 meeting: https://github.com/mpi-forum/mpi-issues/files/4823780/mpi-report-issue120-topol-2020-06-14-annotated.pdf Issue https://github.com/mpi-forum/mpi-issues/issues/120 PR#98 https://github.com/mpi-forum/mpi-standard/pull/98 This is one of the tickets for MPI-4.0 for performance enhancements. It fills the gap that the MPI-1.1 MPI_Dims_create wasn't hardware nor application topology aware. For this, it combines both functionalities, - the factorization of #processes into the given number of dimensions (as non-hardware/application aware in MPI_Dims_create) - the reordering of the processes and providing comm_cart (as in MPI_Cart_create) into one new hardware & application aware routine: MPI_Cart_create_weighted. The hardware awareness comes through the old_comm, the application awareness through the weights. The design of the interface was developed with the help of many forum members (thanks a lot!) and is along the principles we already use for the weights for graph topologies for application topology awareness. This also means, there is enough room for further research and development through an info argument to allow additional features in the future or for specific vendor platforms. The interface is as simple as possible according to the main goal of MPI as defined on page 1 of all MPI standards: "The goal of the Message-Passing Interface simply stated is to develop a widely used standard for writing message-passing programs. As such the interface should establish a practical, portable, efficient, and flexible standard for message passing." For Cartesian applications, the new interface fulfills this short list of being "practical, portable, efficient, and flexible" and can be widely used for all applications that use already the old MPI-1.1 interface MPI_Dims_create + MPI_Cart_create to realize better hardware-awareness and to allow also application-awareness through the new weights. And for portable development of MPI applications, it is equally important that the new interface is part of the MPI standard as it was for the old Cartesian interface. In addition, best optimization may require knowledge of the hardware which may not be disclosed to the public which would prevent third party solutions - this serves as second argument to vote the new interface into the MPI library. Therefore, we'll be glad if you vote for this new performance oriented interface for MPI-4.0. And thank you for reading this announcement to the end. Best regards, Guillaume and the Hardware Topologies Working Group
signature.asc
Description: OpenPGP digital signature
_______________________________________________ mpi-forum mailing list mpi-forum@lists.mpi-forum.org https://lists.mpi-forum.org/mailman/listinfo/mpi-forum
