MPI-3 RMA Implementation Timeline

This page will be used for planning and management of the MPI-3 RMA reference implementation in MPICH2.

The MPICH MPI-3 RMA reference implementation was completed on Nov. 9, 2012.

MPI-3 RMA Reference Implementation Timeline

1. Memory model
   1. Separate -> Unified
      • No changes needed
   2. Accumulate operation ordering
      • No changes needed (CH3 already provides ordering as long as we don't move these operations around in the RMA ops queue).
2. Dynamic windows
   • Done, committed to trunk
3. Shared memory windows
   • Basic support, committed to trunk. Not optimized for shared memory; ops still queue up and will go through CH3 communication.
   • Synchronization operations must be updated for this window type
   1. • Lock - must take the lock immediately
      • Unlock - needs a memory barrier
      • Flush - needs a memory barrier
   • Additional sync operations are needed when flushing shared memory windows.
4. New one-sided communication operations
   1. Compare-and-swap
      • Done, committed to trunk
   2. Fetch-and-op
      • Done, committed to trunk
   3. Get-accumulate
      • Done, committed to trunk.
5. Basic synchronization operations
   1. Ability to acquire lock before unlock
      • Done, committed to trunk.
   2. Ability to issue ops before unlock
      • Done, committed to trunk.
   3. Ability to remove completed operations from the RMA ops list
      • Done, committed to trunk.
   4. Flush
      • Done, committed to trunk.
   5. Sync
      • Done, committed to trunk. Implemented as an OPA_read_write_barrier().
   6. Request-based operations
      • Done, committed to trunk.
6. Multi-target support
   1. Simple locking of multiple targets
      • In CR
   2. Lock-all
      • In CR
   3. Flush-all
      • In CR
7. Info operations: MPI_Win_set_info and MPI_Win_get_info
   1. • Done, committed to trunk.