ProgressThreads

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.

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

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.)

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/.