Jeff Squyres wrote:
...
Two questions then...
1. If the request has already completed, does it mean that since
opal_progress() is not called, no further progress is made?
Correct. It's a latency thing; if your request has already completed,
we just tell you without further delay (i.e., without invoking
opal_progress(), which may trigger lots of other things, and therefore
increase the latency of MPI_REQUEST_GET_STATUS returning).
opal_progress() is our lowest-level progression engine call. It kicks
all kinds of registered progression callbacks from all over the code
base.
2. request->req_complete is tested before calling opal_progress(). Is
it possible that request->req_complete is now true after calling
opal_progress() when this function returns false in *flag?
Yes. I suppose it could be an optimization to duplicate the block
testing for request->req_complete==true below the call to
opal_progress(). I'm guessing the only reason it wasn't done was to
avoid code duplication. Additionally, the call to opal_progress() is
surrounded by an #if block testing OPAL_ENABLE_PROGRESS_THREADS -- if
we have progress threads enabled, the thought was that opal_progress()
(and friends) would be invoked automatically (and probably
continuously) by other threads. The progression thread code is not
well tested -- I'd be surprised if it worked at all, because I doubt
anyone is testing it -- but it has been in our design since the very
beginning. This is likely another reason we don't test again for
req_complete==true after the call to opal_progress() -- because that
block would need to be protected by that #if, leading to further code
complexity.
Hi Jeff,
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)
}
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.
* 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.
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.
Cheers,
Shaun
int i;
for (i = 0; i < 2; i++) {
if( request->req_complete ) {
*flag = true;
/* If this is a generalized request, we *always* have
to call
the query function to get the status (MPI-2:8.2),
even if
the user passed STATUS_IGNORE. */
if (OMPI_REQUEST_GEN == request->req_type) {
ompi_grequest_invoke_query(request,
&request->req_status);
}
if (MPI_STATUS_IGNORE != status) {
*status = request->req_status;
}
return MPI_SUCCESS;
}
#if OMPI_ENABLE_PROGRESS_THREADS == 0
if (i == 0)
opal_progress();
#else
break;
#endif
}
*flag = false;
return MPI_SUCCESS;
