A new release candidate of MPICH, 4.3.0rc3, is now available for download. Since rc2, we have enhanced error reporting in ch4:ofi to include NIC information, fixed a bug in MPI_ANY_SOURCE handling with certain libfabric providers, and reverted a change that caused large messages with ch4:ofi to fallback to MPICH active message protocol. See the CHANGES below for a detailed list of features. At this stage, only critical bugs will be fixed before the 4.3.0 GA release. Please give it a try! You can find the release on our downloads page (https://urldefense.us/v3/__https://www.mpich.org/downloads/__;!!G_uCfscf7eWS... ) Regards, The MPICH team =============================================================================== Changes in 4.3 # Support MPI memory allocation kinds side document. # Support MPI ABI Proposal. Configure with --enable-mpi-abi and build with mpicc_abi. By default, mpicc still builds and links with MPICH ABI. # Experimental API MPIX_Op_create_x. It supports user callback function with extra_state context and op destructor callback. It supports language bindings to use proxy function for language-specific user callbacks. # Experimental API MPIX_{Comm,File,Session,Win}_create_errhandler_x. They allow user error handlers to have extra_state context and corresponding destructor. This allows language bindings to implement user error handlers via proxy. # Experimental API MPIX_Request_is_complete. This is a pure request state query function that will not invoke progress, nor will free the request. This should help applications that want separate task dependency checking from progress engine to avoid progress contentions, especially in multi-threaded context. It is also useful for tools to profile non-deterministic calls such as MPI_Test. # Experimental API MPIX_Async_start. This function let applications to inject progress hooks to MPI progress. It allows application to implement custom asynchronous operations that will be progressed by MPI. It avoids having to implement separate progress mechanisms that may either take additional resource or contend with MPI progress and negatively impact performance. It also allows applications to create custom MPI operations, such as MPI nonblocking collectives, and achieve near native performance. # Added benchmark tests test/mpi/bench/p2p_{latency,bw}. # Added CMA support in CH4 IPC. # Added IPC read algorithm for intranode Allgather and Allgatherv. # Added CVAR MPIR_CVAR_CH4_SHM_POSIX_TOPO_ENABLE to enable non-temporal memcpy for inter-numa shm communication. # Added CVAR MPIR_CVAR_DEBUG_PROGRESS_TIMEOUT for debugging MPI deadlock issues. # ch4:ucx now supports dynamic processes. MPI_Comm_spawn{_multiple} will work. MPI_Open_port will fail due to ucx port name exceeds current MPI_MAX_PORT_NAME of 256. One can work around by use an info hint "port_name_size" and use a larger port name buffer. # PMI-1 defines PMI_MAX_PORT_NAME, which may be different from MPI_MAX_PORT_NAME. This is used by "PMI_Lookup_name". Consequently, MPI_Lookup_name accepts info hint "port_name_size" that may be larger than MPI_MAX_PORT_NAME. If the port name does not fit in "port_name_size", it will return a truncation error. # Autogen default to use -yaksa-depth=2. # Default MPIR_CVAR_CH4_ROOTS_ONLY_PMI to on. # Added ch4 netmod API am_tag_send and am_tag_recv. # Added MPIR_CVAR_CH4_OFI_EAGER_THRESHOLD to force RNDV send mode. # Make check target will run ROMIO tests. # Add back handle conversion macros (f2c/c2f) to preserve ABI compatibility with older MPICH libraries # Fix compilation issue with g++ in -std=gnu++20 mode # Fix bug in MPI_ANY_SOURCE handling observed using the libfabric CXI provider # Add NIC information to error messages in ch4:ofi netmod