[mpich2-dev] Does MPI_WIn_unlock() need a reply?

[David Gingold]

MPICH2 developers --

I've been looking at the ch3 implementation of MPI_Win_unlock().  It  
makes some effort to determine whether it needs to see a message from  
the target indicating that the RMA is complete on that end.  In one  
case, the code concludes that if the lock operation was exclusive  
then there is no need to wait for a reply.

My MPI book says, of MPI_Win_lock(), "RMA operations issued . . .  
will have completed both at the origin and at the target when the  
call returns."

I think I see the argument that the MPICH2 optimization is okay -- if  
it's an exclusive lock, nobody else will be able to get a lock before  
the unlock completes at the target.  But what about something like this:

1.  A locks a window at B, exclusively.
2.  A initiates a put to the window at B.  The put completes at A,  
but not yet at B.
3.  A unlocks a window at B (but waits for no reply)
4.  A indicates to C that A completed its access to B's window
5.  C indicates to B that A has finished A's access
6.  B access its window's memory, but before the put completed at its  
end.

It's a bit of a stretch, yes.  Is it plausible?

-dg

--
David Gingold
Principal Software Engineer
SiCortex
Three Clock Tower Place, Suite 210
Maynard MA 01754
(978) 897-0214 x224



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.mcs.anl.gov/mailman/private/mpich2-dev/attachments/20080303/13a00b8a/attachment.htm>