On Feb 22, 2014, at 10:14 AM, Suraj Prabhakaran <suraj.prabhaka...@gmail.com> wrote:
>> Yeah, we added those capabilities specifically for this purpose. Indeed, >> another researcher added this to Torque a couple of years ago, though it >> didn't get pushed upstream. Also was added to Slurm. > > Thanks for your help . By any chance you have more info on that one? Or a > faint idea where I can find some info on that? I never found something like > that.... The Torque work was being done by someone in Ohio back in 2007 - never heard how it ended up. The name was Prakash Velayutham - you can search the user archives to find the email thread. This was a much older version of OMPI and things have changed a lot. IIRC, someone else since then also asked some questions indicating they were investigating it, but I don't offhand recall who and didn't immediately see it in the archives. The Slurm work was done under my direction while at EMC. Jimmy Cao did the coding, but you won't find anything in our archives about it. It was committed to Slurm about a year ago and is in the current release. You have to configure Slurm to use it, but the orte/mca/ras framework was updated to support a dynamic alloc request (only the Slurm plugin actually implements it today). HTH Ralph > > Best, > Suraj > > On Feb 22, 2014, at 6:38 PM, Ralph Castain wrote: > >> >> On Feb 22, 2014, at 9:30 AM, Suraj Prabhakaran <suraj.prabhaka...@gmail.com> >> wrote: >> >>> Thanks Ralph. >>> >>> I cannot get rid of Torque since I am actually working on dynamic >>> allocation of nodes for a running job on Torque. What I actually want to do >>> is spawn processes on the dynamically assigned nodes since that is the most >>> easiest way to expand MPI processes when a resource allocation is expanded. >> >> No problem - just set "-mca plm rsh" on your cmd line. You'll still use the >> Torque allocator, but use ssh to launch the results. >> >>> >>> I also contemplated on whether any of my changes to the Torque daemons >>> could be the problem but it cannot be because of 2 reasons. >>> >>> 1. For the cases which I have sent you, no dynamic allocation is done. Just >>> using MPI_Comm_spawn on a normal allocation of resources. So my changes to >>> torque are irrelevant here as they are not even called. >>> 2. Further, the processes start successfully on all the nodes. Torque logs >>> don't report any problems and the processes do exist on all the nodes. And >>> the fact that "sometimes" they work and don't have a problem! >>> >>> I am not sure how many users really use MPI_Comm_spawn (spawning large >>> processes) under the Torque environment to actually not such a problem. >>> Because, mpiexec works just fine for any number of processes. >> >> Not many - comm_spawn is only used by researchers upon occasion. I haven't >> seen a "real" application yet, though we may just have not heard about it. >> Still, we do have users with Torque, and perhaps someone can check it. >> >>> >>> Any suggestions or hints on this would be highly appreciated. OpenMPI also >>> seems to be the only implementation we can use for this work at the moment >>> because of the "add-host" info argument for MPI_Comm_spawn which we are >>> using comfortably when spawning onto dynamically allocated hosts which were >>> not a part of the original allocation. >> >> Yeah, we added those capabilities specifically for this purpose. Indeed, >> another researcher added this to Torque a couple of years ago, though it >> didn't get pushed upstream. Also was added to Slurm. >> >> Sadly, I no longer have access to a Torque machine and so I can only offer >> advice. OMPI is executing a state machine, so you could look at one of the >> procs on a machine where they are stalled (look for someone not reporting >> out of the modex) and see where it hung. You can also watch it move thru the >> state machine by setting >> >> -mca state_base_verbose 10 >> >> on your command line. >> >> Happy to provide advice - sorry for the problem >> Ralph >> >> >>> >>> Best, >>> Suraj >>> >>> >>> On Feb 22, 2014, at 4:30 PM, Ralph Castain wrote: >>> >>>> >>>> On Feb 21, 2014, at 5:55 PM, Suraj Prabhakaran >>>> <suraj.prabhaka...@gmail.com> wrote: >>>> >>>>> Hmm.. but in actual the MPI_Comm_spawn of parents and MPI_Init of >>>>> children never returned! >>>> >>>> Understood - my point was that the output shows no errors or issues. For >>>> some reason, the progress thread appears to just stop. This usually >>>> indicates some kind of recursive behavior, but that isn't showing up in >>>> the output. >>>> >>>>> >>>>> I configured MPI with >>>>> >>>>> ./configure --prefix=/dir/ --enable-debug --with-tm=/usr/local/ >>>> >>>> Should be fine. I don't have access to a Torque-based system, and we >>>> aren't hearing issues from other Torque users, so this may have something >>>> to do with how Torque is configured on your system. Perhaps someone with a >>>> Torque-based system on the list could also test this? >>>> >>>> Meantime, I would suggest just using rsh/ssh (since you said that works) >>>> for now as Torque really isn't doing anything for you in this use-case. >>>> >>>> >>>>> >>>>> >>>>> On Feb 22, 2014, at 12:53 AM, Ralph Castain wrote: >>>>> >>>>>> Strange - it all looks just fine. How was OMPI configured? >>>>>> >>>>>> On Feb 21, 2014, at 3:31 PM, Suraj Prabhakaran >>>>>> <suraj.prabhaka...@gmail.com> wrote: >>>>>> >>>>>>> Ok, I figured out that it was not a problem with the node grsacc04 >>>>>>> because I now conducted the same on totally different set of nodes. >>>>>>> >>>>>>> I must really say that with --bind-to none option, the program >>>>>>> completed "many" times compared to earlier but still "sometimes" it >>>>>>> hangs! Attaching now the output of the same case conducted on different >>>>>>> set of nodes with the --bind-to none option. >>>>>>> >>>>>>> mpiexec -mca plm_base_verbose 5 -mca ess_base_verbose 5 -mca >>>>>>> grpcomm_base_verbose 5 --bind-to none -np 3 ./example >>>>>>> >>>>>>> Best, >>>>>>> Suraj >>>>>>> >>>>>>> <output.rtf> >>>>>>> >>>>>>> >>>>>>> On Feb 21, 2014, at 5:03 PM, Ralph Castain wrote: >>>>>>> >>>>>>>> Well, that all looks fine. However, I note that the procs on grsacc04 >>>>>>>> all stopped making progress at the same point, which is why the job >>>>>>>> hung. All the procs on the other nodes were just fine. >>>>>>>> >>>>>>>> So let's try a couple of things: >>>>>>>> >>>>>>>> 1. add "--bind-to none" to your cmd line so we avoid any contention >>>>>>>> issues >>>>>>>> >>>>>>>> 2. if possible, remove grsacc04 from the allocation (you can just use >>>>>>>> the -host option on the cmd line to ignore it), and/or replace that >>>>>>>> host with another one. Let's see if the problem has something to do >>>>>>>> with that specific node. >>>>>>>> >>>>>>>> >>>>>>>> On Feb 21, 2014, at 4:08 AM, Suraj Prabhakaran >>>>>>>> <suraj.prabhaka...@gmail.com> wrote: >>>>>>>> >>>>>>>>> Right, so I have the output here. Same case, >>>>>>>>> >>>>>>>>> mpiexec -mca plm_base_verbose 5 -mca ess_base_verbose 5 -mca >>>>>>>>> grpcomm_base_verbose 5 -np 3 ./simple_spawn >>>>>>>>> >>>>>>>>> Output attached. >>>>>>>>> >>>>>>>>> Best, >>>>>>>>> Suraj >>>>>>>>> >>>>>>>>> <output> >>>>>>>>> >>>>>>>>> On Feb 21, 2014, at 5:30 AM, Ralph Castain wrote: >>>>>>>>> >>>>>>>>>> >>>>>>>>>> On Feb 20, 2014, at 7:05 PM, Suraj Prabhakaran >>>>>>>>>> <suraj.prabhaka...@gmail.com> wrote: >>>>>>>>>> >>>>>>>>>>> Thanks Ralph! >>>>>>>>>>> >>>>>>>>>>> I must have mentioned though. Without the Torque environment, >>>>>>>>>>> spawning with ssh works ok. But Under the torque environment, not. >>>>>>>>>> >>>>>>>>>> Ah, no - you forgot to mention that point. >>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> I started the simple_spawn with 3 processes and spawned 9 processes >>>>>>>>>>> (3 per node on 3 nodes). >>>>>>>>>>> >>>>>>>>>>> There is no problem with the Torque environment because all the 9 >>>>>>>>>>> processes are started on the respective nodes. But the >>>>>>>>>>> MPI_Comm_spawn of the parent and MPI_Init of the children, >>>>>>>>>>> "sometimes" don't return! >>>>>>>>>> >>>>>>>>>> Seems odd - the launch environment has nothing to do with MPI_Init, >>>>>>>>>> so if the processes are indeed being started, they should run. One >>>>>>>>>> possibility is that they aren't correctly getting some wireup info. >>>>>>>>>> >>>>>>>>>> Can you configure OMPI --enable-debug and then rerun the example >>>>>>>>>> with "-mca plm_base_verbose 5 -mca ess_base_verbose 5 -mca >>>>>>>>>> grpcomm_base_verbose 5" on the command line? >>>>>>>>>> >>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> This is the output of simple_spawn - which confirms the above >>>>>>>>>>> statement. >>>>>>>>>>> >>>>>>>>>>> [pid 31208] starting up! >>>>>>>>>>> [pid 31209] starting up! >>>>>>>>>>> [pid 31210] starting up! >>>>>>>>>>> 0 completed MPI_Init >>>>>>>>>>> Parent [pid 31208] about to spawn! >>>>>>>>>>> 1 completed MPI_Init >>>>>>>>>>> Parent [pid 31209] about to spawn! >>>>>>>>>>> 2 completed MPI_Init >>>>>>>>>>> Parent [pid 31210] about to spawn! >>>>>>>>>>> [pid 28630] starting up! >>>>>>>>>>> [pid 28631] starting up! >>>>>>>>>>> [pid 9846] starting up! >>>>>>>>>>> [pid 9847] starting up! >>>>>>>>>>> [pid 9848] starting up! >>>>>>>>>>> [pid 6363] starting up! >>>>>>>>>>> [pid 6361] starting up! >>>>>>>>>>> [pid 6362] starting up! >>>>>>>>>>> [pid 28632] starting up! >>>>>>>>>>> >>>>>>>>>>> Any hints? >>>>>>>>>>> >>>>>>>>>>> Best, >>>>>>>>>>> Suraj >>>>>>>>>>> >>>>>>>>>>> On Feb 21, 2014, at 3:44 AM, Ralph Castain wrote: >>>>>>>>>>> >>>>>>>>>>>> Hmmm...I don't see anything immediately glaring. What do you mean >>>>>>>>>>>> by "doesn't work"? Is there some specific behavior you see? >>>>>>>>>>>> >>>>>>>>>>>> You might try the attached program. It's a simple spawn test we >>>>>>>>>>>> use - 1.7.4 seems happy with it. >>>>>>>>>>>> >>>>>>>>>>>> <simple_spawn.c> >>>>>>>>>>>> >>>>>>>>>>>> On Feb 20, 2014, at 10:14 AM, Suraj Prabhakaran >>>>>>>>>>>> <suraj.prabhaka...@gmail.com> wrote: >>>>>>>>>>>> >>>>>>>>>>>>> I am using 1.7.4! >>>>>>>>>>>>> >>>>>>>>>>>>> On Feb 20, 2014, at 7:00 PM, Ralph Castain wrote: >>>>>>>>>>>>> >>>>>>>>>>>>>> What OMPI version are you using? >>>>>>>>>>>>>> >>>>>>>>>>>>>> On Feb 20, 2014, at 7:56 AM, Suraj Prabhakaran >>>>>>>>>>>>>> <suraj.prabhaka...@gmail.com> wrote: >>>>>>>>>>>>>> >>>>>>>>>>>>>>> Hello! >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> I am having problem using MPI_Comm_spawn under torque. It >>>>>>>>>>>>>>> doesn't work when spawning more than 12 processes on various >>>>>>>>>>>>>>> nodes. To be more precise, "sometimes" it works, and >>>>>>>>>>>>>>> "sometimes" it doesn't! >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> Here is my case. I obtain 5 nodes, 3 cores per node and my >>>>>>>>>>>>>>> $PBS_NODEFILE looks like below. >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> node1 >>>>>>>>>>>>>>> node1 >>>>>>>>>>>>>>> node1 >>>>>>>>>>>>>>> node2 >>>>>>>>>>>>>>> node2 >>>>>>>>>>>>>>> node2 >>>>>>>>>>>>>>> node3 >>>>>>>>>>>>>>> node3 >>>>>>>>>>>>>>> node3 >>>>>>>>>>>>>>> node4 >>>>>>>>>>>>>>> node4 >>>>>>>>>>>>>>> node4 >>>>>>>>>>>>>>> node5 >>>>>>>>>>>>>>> node5 >>>>>>>>>>>>>>> node5 >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> I started a hello program (which just spawns itself and of >>>>>>>>>>>>>>> course, the children don't spawn), with >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> mpiexec -np 3 ./hello >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> Spawning 3 more processes (on node 2) - works! >>>>>>>>>>>>>>> spawning 6 more processes (node 2 and 3) - works! >>>>>>>>>>>>>>> spawning 9 processes (node 2,3,4) - "sometimes" OK, "sometimes" >>>>>>>>>>>>>>> not! >>>>>>>>>>>>>>> spawning 12 processes (node 2,3,4,5) - "mostly" not! >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> I ideally want to spawn about 32 processes with large number of >>>>>>>>>>>>>>> nodes, but this is at the moment impossible. I have attached my >>>>>>>>>>>>>>> hello program to this email. >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> I will be happy to provide any more info or verbose outputs if >>>>>>>>>>>>>>> you could please tell me what exactly you would like to see. >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> Best, >>>>>>>>>>>>>>> Suraj >>>>>>>>>>>>>>> >>>>>>>>>>>>>>> <hello.c>_______________________________________________ >>>>>>>>>>>>>>> devel mailing list >>>>>>>>>>>>>>> de...@open-mpi.org >>>>>>>>>>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>>>>>>>>>>>>> >>>>>>>>>>>>>> _______________________________________________ >>>>>>>>>>>>>> devel mailing list >>>>>>>>>>>>>> de...@open-mpi.org >>>>>>>>>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>>>>>>>>>>>> >>>>>>>>>>>>> _______________________________________________ >>>>>>>>>>>>> devel mailing list >>>>>>>>>>>>> de...@open-mpi.org >>>>>>>>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>>>>>>>>>>> >>>>>>>>>>>> _______________________________________________ >>>>>>>>>>>> devel mailing list >>>>>>>>>>>> de...@open-mpi.org >>>>>>>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>>>>>>>>>> >>>>>>>>>>> _______________________________________________ >>>>>>>>>>> devel mailing list >>>>>>>>>>> de...@open-mpi.org >>>>>>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>>>>>>>>> >>>>>>>>>> _______________________________________________ >>>>>>>>>> devel mailing list >>>>>>>>>> de...@open-mpi.org >>>>>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>>>>>>>> >>>>>>>>> _______________________________________________ >>>>>>>>> devel mailing list >>>>>>>>> de...@open-mpi.org >>>>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>>>>>>> >>>>>>>> _______________________________________________ >>>>>>>> devel mailing list >>>>>>>> de...@open-mpi.org >>>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>>>>>> >>>>>>> _______________________________________________ >>>>>>> devel mailing list >>>>>>> de...@open-mpi.org >>>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>>>>> >>>>>> _______________________________________________ >>>>>> devel mailing list >>>>>> de...@open-mpi.org >>>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>>>> >>>>> _______________________________________________ >>>>> devel mailing list >>>>> de...@open-mpi.org >>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>>> >>>> _______________________________________________ >>>> devel mailing list >>>> de...@open-mpi.org >>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >>> >>> _______________________________________________ >>> devel mailing list >>> de...@open-mpi.org >>> http://www.open-mpi.org/mailman/listinfo.cgi/devel >> >> _______________________________________________ >> devel mailing list >> de...@open-mpi.org >> http://www.open-mpi.org/mailman/listinfo.cgi/devel > > _______________________________________________ > devel mailing list > de...@open-mpi.org > http://www.open-mpi.org/mailman/listinfo.cgi/devel
