Implementing a Progress Thread approach in Open MPI

At the spring 2011 developers meeting, it was decided that Open MPI needed to support progress threads, for a number of reasons:

• progression of long messages is poor with the existing implementation
• non-blocking collectives really need async progress to make sense in many HPC apps
• the MPI-3 one-sided makes significantly stronger statements about progress than MPI-2, and we can't currently meet MPI-2's requirements

This page documents the plan and status of implementing the progress threads in Open MPI.

People

• Brian Barrett
• Ralph Castain

Please add yourself to the list if you're willing to participate.

Project Plan

• Create branch for progress threads work (DONE)
• Create nightly regression test runs of progress threads and get people to run them
• Migrate away from central progress engine to per-project progress
  • Each project (opal/orte/ompi) will be responsible for progress, but opal will still provide the event library for building those progress engines. It is possible we will reintegrate if that proves useful (or have a threaded progress engine for all to use with events or...).
• Shame component authors to work on the thread support in their components
• Merge into trunk after 1.7 branch (or before if we're done already. Ha.)

Implementation Notes

Current Status

There is a branch which tracks the trunk (when Brian remembers to update it) at https://svn.open-mpi.org/svn/ompi/tmp-public/progress-threads/.