To: J3 J3/19-263 From: Bill Long & Lorri Menard Subject: report for MPI Date: 2019-October-24 The current goal is to have MPI-4 done in time for Super Computing 2020. Fault tolerance -------------------- It looks like some basic fault tolerance functionality will be making it into the MPI 4.0 standard in 2020. Non-controversial parts of ULFM are being adopted. The forum is still considering ULFM and the Re-init proposal. (ULFM - User Level Failure Mitigation). MPI_Cart_create_weighted ------------------------------------- Rolf Rabenseifner proposed a new cartesian communicator function called MPI_Cart_create_weighted. The motivation is to improve performance. User supplied weights to indicate the estimated communication intensity for eacn dimension of an array being spread across the cartesian domain. Big count ------------- The "big count" proposal is expected to make it into the MPI 4.0 standard. There's still plenty of discussion about how exactly to implement it. Partitioned communication ------------------------------------- This is a rebranding of a previous proposal known as "fine points". The point of "partitioned communication" is to provide a simple way for multiple threads to participate in network operations. In the proposal, a buffer is partitioned into chunks and different threads can handle different chunks of the buffer. The current reincarnation of this proposal suggests the use of a persistent communication API, whereas the previous proposal was based on plain old send's and recv's. 1. MPI_T events - https://www.mpi-forum.org/bofs/2018-11-sc/events.pdf 2. MPI sessions - https://www.mpi-forum.org/bofs/2018-11-sc/sessions.pdf 3. Support for 16-bit floating point operations, e.g., for MPI_Allreduce operations used to implement neural networks