I haven't worked on rma before but was
working on a problem and ran into this **comment in MPI_Win_post:
"Starts an RMA exposure epoch for
the local window associated with win. **Only the processes belonging
to group should access the window with RMA calls on win during this epoch.
Each process in group must issue a matching call to MPI_Win_start.
MPI_Win_post does not block."
Would overlapping epochs be violating
the ** line? I decided I probably need to support this but I wondered
if it's bending or breaking the 'rules'?
The problem (code at the bottom of this
email) is using a cartesian communicator and alternating "left/right'
accumulates with 'up/down' accumulates on a single win. So:
- Ranks 0,1,2,3 are doing a left/right
accumulate.
- Ranks 4,5,6,7 are doing a left/right
accumulate.
- ...
and then sometimes...
- Ranks 0,1,2,3 complete and enter the
'up/down' accumulate epoch
-- Rank 0 does MPI_Win_post to ranks
4,12
-- Rank 1 doesn MPI_Win_post to ranks
5,13
...
So is Rank 0 posting to Rank 4 while
4 is still in the epoch with 5/6/7 a violation of "Only the processes
belonging to group should access the window with RMA calls on win during
this epoch"? From Rank 4's point of view, rank 0 isn't in the
group for the current win/epoch.
Putting a barrier (or something) in
between or using two different win's fixes it. I like using two win's
since it separates the epochs and clearly doesn't use the wrong group/rank
on the win.
/* RMA transfers in left-right
direction */
MPI_Win_post(grp_lr, 0,
win);
MPI_Win_start(grp_lr,
0, win);
MPI_Accumulate(&i,
1, MPI_INT, ranks_lr[LEFT] , 0, 1, MPI_INT, MPI_SUM, win);
MPI_Accumulate(&i,
1, MPI_INT, ranks_lr[RIGHT], 0, 1, MPI_INT, MPI_SUM, win);
MPI_Win_complete(win);
MPI_Win_wait(win);
/* RMA transfers in up-down
direction */
MPI_Win_post(grp_ud, 0,
win);
MPI_Win_start(grp_ud,
0, win);
MPI_Accumulate(&i,
1, MPI_INT, ranks_ud[UP] , 0, 1, MPI_INT, MPI_SUM, win);
MPI_Accumulate(&i,
1, MPI_INT, ranks_ud[DOWN], 0, 1, MPI_INT, MPI_SUM, win);
MPI_Win_complete(win);
MPI_Win_wait(win);