Yes, the code is restartable, and our users often do this. We have users in countries with unreliable power supplies. However, we still try to make the code as robust as possible. Usually, if I do something improper in my MPI coding, failure occurs right away. But I've run out of ideas as to why the code would fail after two days, other than network hiccups beyond my control. In this case, I run 15 jobs simultaneously on a new linux cluster that is dedicated solely to our program. Of the 15 jobs, maybe none or one will fail. So that makes me think there is something random at play, but I can't think of ways to further bullet-proof the code.
When you say "poorly designed I/O", what do you mean? This is something that I have not really considered. -----Original Message----- From: users [mailto:users-boun...@open-mpi.org] On Behalf Of Gus Correa Sent: Thursday, October 16, 2014 11:40 AM To: Open MPI Users Subject: Re: [OMPI users] Hybrid OpenMPI/OpenMP leading to deadlocks? Hi Kevin Wouldn't it be possible to make your code restartable, by saving the appropriate fluid configuration/phase space variables, and splitting your long run into smaller pieces? That is a very common strategy for large PDE integrations. Time invested in programming the restart features may pay off in more steady jobs, and in integrations that actually complete. If a long run fails, because of a network glitch, exhaustion of resources, slow NFS server, or any other reason, you loose a long computing time. If a short run fails, you can always restart from the previous leg/job, and your loss is small. Nearly all parallel code that we run here is restartable (GFD is a form of CFD :) ), both to prevent long run failures like the one you mentioned, as well as to provide a better throughput and user time share on the cluster as a whole. This is nothing new, and restartable codes + short job queue time policy is very common out there. Here most problems with long runs (we have some non-restartable serial code die-hards), happen due to NFS issues (busy, slow response, etc), and code with poorly designed IO. My two cents, Gus Correa On 10/16/2014 10:16 AM, McGrattan, Kevin B. Dr. wrote: > The individual MPI processes appear to be using a few percent of the > system memory. > > I have created a loop containing repeated calls to MPI_TESTALL. When > the process is in this loop for more than 10 s, it calls MPI_ABORT. So > the only error message I see is related to all the processes being > stopped suddenly. > > Is it reasonable to ABORT after 10 s? If I just use MPI_WAITALL, the > job hangs indefinitely. I at least know for which MPI exchange and > which MPI process the code is hanging. > > Is there a way to change the number of retries for a given MPI > send/receive, or a more graceful timeout function than what I have coded? > > *From:*users [mailto:users-boun...@open-mpi.org] *On Behalf Of *Ralph > Castain > *Sent:* Wednesday, October 15, 2014 11:05 PM > *To:* Open MPI Users > *Subject:* Re: [OMPI users] Hybrid OpenMPI/OpenMP leading to deadlocks? > > If you only have one thread doing MPI calls, then single and funneled > are indeed the same. If this is only happening after long run times, > I'd suspect resource exhaustion. You might check your memory footprint > to see if you are running into leak issues (could be in our library as > well as your app). When you eventually deadlock, do you get any error output? > If you are using IB and run out of QP, you should at least get > something saying that. > > On Oct 15, 2014, at 8:22 AM, McGrattan, Kevin B. Dr. > <kevin.mcgrat...@nist.gov <mailto:kevin.mcgrat...@nist.gov>> wrote: > > > > I am using OpenMPI 1.8.3 on a linux cluster to run fairly long CFD > (computational fluid dynamics) simulations using 16 MPI processes. The > calculations last several days and typically involve millions of MPI > exchanges. I use the Intel Fortran compiler, and when I compile with > the -openmp option and run with only one OpenMP thread per MPI > process, I tend to get deadlocks after several days of computing. > These deadlocks only occur in about 1 out of 10 calculations, and they > only occur after running for several days. I have eliminated things > like network glitches, power spikes, etc, as possibilities. The only > thing left is the inclusion of the OpenMP option - even though I am > running with just one OpenMP thread per MPI process. I have read about > the issues with MPI_THREAD_INIT, and I have reduced the REQUIRED level > of support to MPI_THREAD_FUNNELED, down from MPI_THREAD_SERIALIZED. > The latter was not necessary for my application, and I think the > reduction in level of support has helped, but not completely removed, > the deadlocking. Of course, there is always the possibility that I > have coded my MPI calls improperly, even though the code runs for days > on end. Maybe there's that one in a million possibility that rank x > gets to a point in the code that is so far ahead of all the other ranks that > a deadlock occurs. > Placing MPI_BARRIERs has not helped me find any such situation. > > So I have two questions. First, has anyone experienced something > similar to this where inclusion of OpenMP in an MPI code has caused deadlocks? > Second, is it possible that reducing the REQUIRED level of support to > MPI_THREAD_SINGLE will cause the code to behave differently than > FUNNELED? I have read in another post that SINGLE and FUNNELED are > essentially the same thing. I have even noted that I can spawn OpenMP > threads even when I use SINGLE. > > Thanks > > Kevin McGrattan > > National Institute of Standards and Technology > > 100 Bureau Drive, Mail Stop 8664 > > Gaithersburg, Maryland 20899 > > 301 975 2712 > > _______________________________________________ > users mailing list > us...@open-mpi.org <mailto:us...@open-mpi.org> > Subscription:http://www.open-mpi.org/mailman/listinfo.cgi/users > Link to this > post:http://www.open-mpi.org/community/lists/users/2014/10/25500.php > > > > _______________________________________________ > users mailing list > us...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users > Link to this post: > http://www.open-mpi.org/community/lists/users/2014/10/25506.php > _______________________________________________ users mailing list us...@open-mpi.org Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/users Link to this post: http://www.open-mpi.org/community/lists/users/2014/10/25507.php
