Cisco Blogs

MPI-3.1: It’s official!

June 4, 2015 - 0 Comments

Woo hoo!

MPI 3.1 votes: aside from one organization who had the leave before the final vote, the MPI Forum unanimously voted to approve MPI-3.1.

Get your fresh, hot copy of the MPI-3.1 specification (I’m told that physical books will become available in the coming months for those who are interested — can you guess what color the books will be?).

MPI-3.1 is mainly a bug fix release.  Here’s a brief list of what has changed since MPI-3.0:

  • Added address addition/subtraction operators (MPI_AINT_ADD / MPI_AINT_DIFF)
  • Defined several thread query functions to always be thread safe (e.g., MPI_INITIALIZED)
  • Added same_disp_unit info key for RMA operations
  • Add several clarifications about the MPI_T tools interface
  • Various fixes for the mpi_f08 bindings:
    • Fixed all the problems regarding the use of BIND(C)
    • Specify the types of TYPE(MPI_Status) derived type members
    • Fix the flag argument type for MPI_IMPROBE
    • Fix the newcomm argument in MPI_COMM_IDUP
    • Fix the comm and ierr arguments in MPI_COMM_SET_INFO
    • Fix resultlen argument in MPI_GET_LIBRARY_VERSION
    • Fixed definitions and specifications of the PMPI profiling interface
  • Clarify uses of virtual graph topologies in neighborhood collectives
  • Clarify usage of same_size info key with windows
  • Clarify use of MPI_NO_OP and other roles issues with MPI_GET_ACCUMULATE
  • Added missing MPI_CONVERSION_FN_NULL in annexes

Brought to you by:


– The MPI 3.1 Forum

In an effort to keep conversations fresh, Cisco Blogs closes comments after 60 days. Please visit the Cisco Blogs hub page for the latest content.