[mpi3-ft] FW: Starting slides for 2/1/2008 telecon

Thomas Herault thomas.herault at lri.fr
Wed Jan 30 03:40:10 CST 2008

Le 30 janv. 08 à 09:17, Greg Bronevetsky a écrit :

> Two comments:
> We may want to add the capability to spawn new processes and give
> them the ranks of the failed processes. This is more efficient than
> pre-allocating enough spare processes as part of the original job
> allocation, so it might be a good idea to include in the spec.

A user should be able to spawn new processes and create a new
communicator with enough processes to circumvent this issue.

I think that size of MPI_COMM_WORLD should decrease
when failures occur. Not all fault tolerance solutions need processor
replacement (e.g. a master/worker approach when a worker fails).
But MPI_UNIVERSE_SIZE should be allowed to remain constant
(thus, the system can provide dynamically new processors to
replace failed ones).

> I disagree with the comments about MPI quieting the communication
> system because this presumes that the application will use the
> trivial sync-and-stop CPR protocol. They may the case but we
> shouldn't write this assumption into the spec. We should probably
> restrict ourselves to only saying that no message may get partially
> delivered since such messages would be very hard to deal with above
> the MPI library.

I agree that we should not assume that synchronized CPR will be the
only approach used. Messages must certainly be kept transactional
(either completely received or not at all). But what about collective
communications? As suggested during the meeting, a two-phase commit
protocol can be enforced to ensure that any collective communication
either completes or fails on any living processor; however this may be
considered as too inefficient for the normal case, when failures do not

Thomas Herault
assistant professor
INRIA/Univ. Paris Sud

> Greg Bronevetsky
> Post-Doctoral Researcher
> 1028 Building 451
> Lawrence Livermore National Lab
> (925) 424-5756
> bronevetsky1 at llnl.gov
> At 10:12 AM 1/29/2008, Richard Graham wrote:
>> This did not seem to make it through the first time, so let me try  
>> again.
>> Rich
>> ------ Forwarded Message
>> From: Richard Graham <rlgraham at ornl.gov>
>> Date: Tue, 29 Jan 2008 10:55:11 -0500
>> To: Discussion of MPI 3 Fault Tolerance Support <mpi3-ft at cs.uiuc.edu>
>> Conversation: Starting slides for 2/1/2008 telecon
>> Subject: Starting slides for 2/1/2008 telecon
>> Attached is a set of slides I intend to use as a staring point for  
>> the
>> telecon this coming Friday.  If you are planning on attending,  
>> please take a
>> look at these, and see what is missing.  The main goal for this  
>> call is to
>> help set the scope of the problem for which we intend to propose  
>> changes to
>> the MPI standard.
>> Thanks,
>> Rich
>> ------ End of Forwarded Message
>> _______________________________________________
>> mpi3-ft mailing list
>> mpi3-ft at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/mpi3-ft
> _______________________________________________
> mpi3-ft mailing list
> mpi3-ft at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/mpi3-ft

More information about the mpiwg-ft mailing list