Skip to content

v5.7
Compare
Choose a tag to compare
@jeffdaily jeffdaily released this 30 Mar 22:41
· 640 commits to master since this release
v5.7
1537981
  • Known Bugs
    • Some combinations of MPI implementations with the MPI RMA and PR
      ports fail. Recommended to use latest MPI implementations available.
  • Added
    • Tiled data layout
    • Read-only property type using replication across SMP nodes
  • Changed

    • GA is now thread safe

    • MPI3 implementation based on MPI RMA now uses data types in MPI
      calls by default. This is higher performing but not as reliable as
      using multiple contiguous data transfers. The build can be
      configured to use contiguous transfers if data types are not working
      for your MPI implementation.

    • ComEx MPI-PR now uses MPI data types in strided put and get calls
      by default. To enable the old packed behavior, set the following
      environment variables to 0.

      • COMEX_ENABLE_PUT_DATATYPE
      • COMEX_ENABLE_GET_DATATYPE

      Additionally, the original packing implementation is faster for smaller
      messages. Two new environment variables control at which point the MPI
      data types are used.

      • COMEX_PUT_DATATYPE_THRESHOLD. Default 8192.
      • COMEX_GET_DATATYPE_THRESHOLD. Default 8192.
  • Fixed
    • Message sizes exceeding 2GB now work correctly
    • Mirrored Arrays now distributes data across SMP nodes for
      ComEx-based runtimes
    • Matrix multiply works for non-standard data layouts (may not be
      performant)
  • Closed Issues
    • [#48] Message sizes exceeding 2GB may not work correctly
Assets 3
Loading