Hi Jeff,
Jeff Squyres wrote:
On Sep 8, 2009, at 1:06 PM, Shaun Jackman wrote:
My INBOX has been a disaster recently. Please ping me repeatedly if
you need quicker replies (sorry! :-( ).
(btw, should this really be on the devel list, not the user list?)
It's tending that way. I'll keep the thread here for now for
continuity. If I start a new thread on this topic, I'll move it to devel.
I can see one sort of ugly scenario unfolding in my head. Consider two
processes running the following pseudocode:
req = MPI_Irecv
while (!done) {
while (MPI_Test(req)) {
req = MPI_Irecv
}
MPI_Send(!me)
MPI_Send(!me)
}
Are the sends guaranteed to have matching receives elsewhere? If not,
this has potential to deadlock on the whole assuming-buffering issue...
You're right that this is an erroneous program because there is only
one Irecv posted and two Send. Change the two MPI_Send to MPI_Bsend to
prevent deadlock, and the situation I'm describing below still applies.
If you're expecting the sends to be matched by the Irecv's, this looks
like an erroneous program to me (there will always be 2x as many sends
outstanding as receives).
I'll describe one process here:
* MPI_Test checks req->req_complete, which is false, then calls
opal_progress (which finds two packets from the other guy).
* Send two packets to the other guy.
...only if they're eager. The sends are *not* guaranteed to complete
until the matching receives occur.
* MPI_Test checks req->req_complete, which is true, returns
immediately. No progress is made.
* MPI_Test checks req->req_complete, which is false, because no
progress has been made since the last call. Call opal_progress (which
finds two packets from the other guy).
* Send two packets to the other guy.
* MPI_Test checks req->req_complete, which is true, returns
immediately. No progress is made.
* MPI_Test checks req->req_complete, which is false, because no
progress has been made since the last time. Call opal_progress (which
finds two packets from the other guy).
* Send two packets to the other guy.
and loop.
In each iteration through the loop, one packet is received and two
packets are sent. Eventually this has to end badly.
Bad user behavior should be punished, yes. :-)
I'm not quite sure that I see the problem you're identifying -- from
what you describe, I think it's an erroneous program.
With buffered sends, in each iteration through the loop two packets
will be sent and only one packet will be received due to
MPI_Request_get_status not checking request->req_complete after
calling opal_progress.
Following is an untested fix to request_get_status.c. It checks
req->req_complete and returns immediately if it is true. If not, it
calls opal_progress() and checks req->req_complete again. If
OMPI_ENABLE_PROGRESS_THREADS is defined, it only checks the once and
does not call opal_progress(). It would look better if the body of the
loop were factored out into its own function.
Hmm. Do you mean this to be in request_get_status.c or req_test.c?
(you mentioned MPI_TEST above, not MPI_REQUEST_GET_STATUS)
I meant this code for MPI_Request_get_status. I just read the code for
ompi_request_default_test in req_test.c. It includes code very similar
to what I suggested for checking request->req_complete, calling
opal_progress and then checking request->req_complete a second time,
except that it implements the loop using a goto.
Is this the optimization I mentioned in my previous reply (i.e., if
req_complete is false, call opal_progress, and then check req_complete
again?) If so, I think it would be better to do it without an if loop
somehow (testing and branching, etc.).
Yes, and MPI_Request_get_status would then behave as MPI_Test does
currently. Is it not so crazy for MPI_Test to be implemented as calls
to MPI_Request_get_status and MPI_Request_free? It would eliminate the
code duplication.
Cheers,
Shaun