If we were to change the MPI_ANY_SOURCE semantics to only return when completed/matched (so the suggested proposal modification), would this make life more difficult on an intermediate layer virtualizing the collective operations? I am open to the idea that calling a collective validate() also re-enables the posting of MPI_ANY_SOURCE. The reason we did not do that in the current proposal is because we were uncertain if someone would want to do a validate, but not re-enable the posting of MPI_ANY_SOURCE receives. Since reenabling ANY_SOURCE is a local/quick operation there was not performance argument to be made. Though I think you pose an interesting programability argument. So can you elaborate a bit on your example (I just want to make sure I fully understand)? Would such an intermediate layer be using MPI_ANY_SOURCE in their collective operations and depend on the return-when-new-proc-failure semantic? So the intermediate library would have to reenable ANY_SOURCE, then mask the semantics for user initiated p2p operations over the same communicator. If we added the reenable_any_source semantics with the validate then the intermediate library would not have to virtualize the p2p communication. Is that on the right track? Thanks, Josh On Wed, Jan 25, 2012 at 7:29 PM, Martin Schulz <[email protected]> wrote:

Hi Josh, all,

I agree, I think the current approach is fine. Blocking is likely to be more problematic in many cases, IMHO. However, I am still a bit worried about splitting the semantics for P2P and Collective routines. I don't see a reason why a communicator after a collective call to validate wouldn't support ANY_SOURCE. If it is split, though, any intermediate layer trying to replace collectives with P2P solutions (and there are plenty of tuning frameworks out there that try exactly that) have a hard time maintaing the same semantics in an error case.

Martin

On Jan 25, 2012, at 3:06 PM, Howard Pritchard wrote:

...

Hi Josh,

Cray is okay with the semantics described in the current FTWG proposal attached to the ticket.

We plan to just leverage the out-of-band system fault detector software that currently kills jobs if a node goes down that the job was running on.

Howard

Josh Hursey wrote:

...

We really need to make a decision on semantics for MPI_ANY_SOURCE.

During the plenary session the MPI Forum had a problem with the current proposed semantics. The current proposal states (roughly) that MPI_ANY_SOURCE return when a failure emerges in the communicator. The MPI Forum read this as a strong requirement for -progress- (something the MPI standard tries to stay away from).

The alternative proposal is that a receive on MPI_ANY_SOURCE will block until completed with a message. This means that it will -not- return when a new failure has been encountered (even if the calling process is the only process left alive in the communicator). This does get around the concern about progress, but puts a large burden on the end user.

There are a couple good use cases for MPI_ANY_SOURCE (grumble, grumble) - Manager/Worker applications, and easy load balancing when multiple incoming messages are expected. This blocking behavior makes the use of MPI_ANY_SOURCE dangerous for fault tolerant applications, and opens up another opportunity for deadlock.

For applications that want to use MPI_ANY_SOURCE and be fault tolerant they will need to build their own failure detector on top of MPI using directed point-to-point messages. A basic implementation might post MPI_Irecv()'s to each worker process with an unused tag, then poll on Testany(). If any of these requests complete in error (MPI_ERR_PROC_FAIL_STOP) then the target has failed and the application can take action. This user-level failure detector can (should) be implemented in a third-party library since failure detectors can be difficult to implement in a scalable manner.

In reality, the MPI library or the runtime system that supports MPI will already be doing something similar. Even for MPI_ERRORS_ARE_FATAL on MPI_COMM_WORLD, the underlying system must detect the process failure, and terminate all other processes in MPI_COMM_WORLD. So this represents a -detection- of the failure, and a -notification- of the failure throughout the system (though the notification is an order to terminate). For MPI_ERRORS_RETURN, the MPI will use this detection/notification functionality to reason about the state of the message traffic in the system. So it seems silly to force the user to duplicate this (nontrivial) detection/notification functionality on top of MPI, just to avoid the progress discussion.

So that is a rough summary of the debate. If we are going to move forward, we need to make a decision on MPI_ANY_SOURCE. I would like to make such a decision before/during the next teleconf (Feb. 1).

I'm torn on this one, so I look forward to your comments.

-- Josh

-- Joshua Hursey Postdoctoral Research Associate Oak Ridge National Laboratory http://users.nccs.gov/~jjhursey

-- Howard Pritchard Software Engineering Cray, Inc. _______________________________________________ mpi3-ft mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft

________________________________________________________________________ Martin Schulz, [email protected], http://people.llnl.gov/schulzm CASC @ Lawrence Livermore National Laboratory, Livermore, USA

_______________________________________________ mpi3-ft mailing list [email protected] http://lists.mpi-forum.org/mailman/listinfo.cgi/mpi3-ft

-- Joshua Hursey Postdoctoral Research Associate Oak Ridge National Laboratory http://users.nccs.gov/~jjhursey