Re: [OMPI users] Restart after code hangs
Hi Alex You know all this, but just in case ... Restartable code goes like this: * .start read the initial/previous configuration from a file ... final_step = first_step + nsteps time_step = first_step while ( time_step .le. final_step ) ... march in time ... time_step = time_step +1 end save the final_step configuration (or phase space) to a file [depending on the algorithm you may need to save the previous config also, or perhaps a few more] .end Then restart the job time and again, until the desired number of time steps is completed. Job queue systems/resource managers allow a job to resubmit itself, so unless a job fails it feels like a single time integration. If a job fails in the middle, you don't lose all work, just restart from the previous saved configuration. That is the only situation where you need to "monitor" the code. Resource managers/ queue systems can also email you in case the job fails, warning you to do manual intervention. The time granularity per job (nsteps) is up to you. Normally it is adjusted to the max walltime of job queues (in a shared computer/cluster), but in your case it can be adjusted to how often the program fails. All atmosphere/ocean/climate/weather_forecast models work this way (that's what we mostly run here). I guess most CFD, computational Chemistry, etc, programs also do. I hope this helps, Gus Correa On 06/16/2016 05:25 PM, Alex Kaiser wrote: Hello, I have an MPI code which sometimes hangs, simply stops running. It is not clear why and it uses many large third party libraries so I do not want to try to fix it. The code is easy to restart, but then it needs to be monitored closely by me, and I'd prefer to do it automatically. Is there a way, within an MPI process, to detect the hang and abort if so? In psuedocode, I would like to do something like for (all time steps) step if (nothing has happened for x minutes) call mpi abort to return control to the shell endif endfor This structure does not work however, as it can no longer do anything, including check itself, when it is stuck. Thank you, Alex ___ users mailing list us...@open-mpi.org Subscription: https://www.open-mpi.org/mailman/listinfo.cgi/users Link to this post: http://www.open-mpi.org/community/lists/users/2016/06/29471.php
Re: [OMPI users] Restart after code hangs
Which version of OMPI are you using? > On Jun 16, 2016, at 2:25 PM, Alex Kaiserwrote: > > Hello, > > I have an MPI code which sometimes hangs, simply stops running. It is not > clear why and it uses many large third party libraries so I do not want to > try to fix it. The code is easy to restart, but then it needs to be monitored > closely by me, and I'd prefer to do it automatically. > > Is there a way, within an MPI process, to detect the hang and abort if so? > > In psuedocode, I would like to do something like > for (all time steps) > step > if (nothing has happened for x minutes) > call mpi abort to return control to the shell > endif > endfor > This structure does not work however, as it can no longer do anything, > including check itself, when it is stuck. > > Thank you, > Alex > > ___ > users mailing list > us...@open-mpi.org > Subscription: https://www.open-mpi.org/mailman/listinfo.cgi/users > Link to this post: > http://www.open-mpi.org/community/lists/users/2016/06/29471.php
[OMPI users] Restart after code hangs
Hello, I have an MPI code which sometimes hangs, simply stops running. It is not clear why and it uses many large third party libraries so I do not want to try to fix it. The code is easy to restart, but then it needs to be monitored closely by me, and I'd prefer to do it automatically. Is there a way, within an MPI process, to detect the hang and abort if so? In psuedocode, I would like to do something like for (all time steps) step if (nothing has happened for x minutes) call mpi abort to return control to the shell endif endfor This structure does not work however, as it can no longer do anything, including check itself, when it is stuck. Thank you, Alex
[OMPI users] Avoiding the memory registration costs by having memory always registered, is it possible with Linux ?
Hi, After reading a little the FAQ on the methods used by Open MPI to deal with memory registration (or pinning) with Infiniband adapter, it seems that we could avoid all the overhead and complexity of memory registration/deregistration, registration cache access and update, memory management (ummunotify) in addition to allowing a better overlap of the communications with the computations (we could let the communication hardware do its job independently without resorting to registration/transfer/deregistration pipelines) by simply having all user process memory registered all the time. Of course a configuration like that is not appropriate in a general setting (ex: a desktop environment) as it would make swapping almost impossible. But in the context of an HPC node where the processes are not supposed to swap and the OS not overcommit memory, not being able to swap doesn't appear to be a problem. Moreover since the maximal total memory used per process is often predefined at the application start as a resource specified to the queuing system, the OS could easily keep a defined amount of extra memory for its own need instead of swapping out user process memory. I guess that specialized (non-Linux) compute node OS does this. But is it possible and does it make sense with Linux ? Thanks, Martin Audet
[OMPI users] "failed to create queue pair" problem, but settings appear OK
Thank-you Nathan. Since the default btl_openib_receive_queues setting is: P,128,256,192,128:S,2048,1024,1008,64:S,12288,1024,1008,64:S,65536,1024,1008,64 this would mean that, with max_qp = 392632 and 4 QPs above, the "actual" max would be 392632 / 4 = 98158. Using this value in my prior math, the upper bound on the number of 24-core nodes would be 98158 / 24^2 ~ 170.This comes closer to the limit I encountered while testing. I'm sure there are other particulars I am not accounting for in this math, but the approximation is reasonable. Thanks for the clarification, Nathan! --john -Original Message- From: users [mailto:users-boun...@open-mpi.org] On Behalf Of Nathan Hjelm Sent: Thursday, June 16, 2016 9:56 AM To: Open MPI Users Subject: EXT: Re: [OMPI users] "failed to create queue pair" problem, but settings appear OK XRC support is greatly improved in 1.10.x and 2.0.0. Would be interesting to see if a newer version fixed the shutdown hang. When calculating the required number of queue pairs you also have to divide by the number of queue pairs in the btl_openib_receive_queues parameter. Additionally Open MPI uses 1 qp/rank for connections (1.7+) and there are some in use by IPoIB and other services. -Nathan > On Jun 16, 2016, at 7:15 AM, Sasso, John (GE Power, Non-GE) >wrote: > > Nathan, > > Thank you for the suggestion. I tried your btl_openib_receive_queues > setting with a 4200+ core IMB job, and the job ran (great!). The shutdown > of the job took such a long time that after 6 minutes, I had to > force-terminate the job. > > When I tried using this scheme before, with the following recommended by the > OpenMPI FAQ, I got some odd errors: > > --mca btl openib,sm,self --mca btl_openib_receive_queues > X,128,256,192,128:X,2048,256,128,32:X,12288,256,128,32:X,65536,256,128 > ,32 > > However, when I tried: > > --mca btl openib,sm,self --mca btl_openib_receive_queues > X,4096,1024:X,12288,512:X,65536,512 > > I got success with my aforementioned job. > > I am going to do more testing, with the goal of getting a 5000 core job to > run successfully. If I can, then down the road my concern is the impact the > btl_openib_receive_queues mca parameter (above) will have on lower-scale (< > 1024 cores) jobs, since the parameter change to the global openmpi config > file would impact ALL users and jobs of all scales. > > Chuck – as I noted in my first email, log_num_mtt was set fine, so that is > not the issue here. > > Finally, regarding running out of QPs, I examined the output of ‘ibv_devinfo > –v’ on our compute nodes. I see the following pertinent settings: > > max_qp: 392632 > max_qp_wr: 16351 > max_qp_rd_atom: 16 > max_qp_init_rd_atom:128 > max_cq: 65408 >max_cqe:4194303 > > Figuring that max_qp is the prime limitation here I am running into when > using the PP and SRQ QPs, considering 12 cores per node, this would seem to > imply that an upper bound on the number of nodes would be 392632 / 24^2 ~ 681 > nodes. This does not make sense, because I saw the QP creation failure error > (again, NO error about failure to register enough memory) for as small as 177 > 24-core nodes! I don’t know how to make sense of this, tho I don’t question > that we were running out of QPs. > > --john > > > From: users [mailto:users-boun...@open-mpi.org] On Behalf Of Nathan > Hjelm > Sent: Wednesday, June 15, 2016 2:43 PM > To: Open MPI Users > Subject: EXT: Re: [OMPI users] "failed to create queue pair" problem, > but settings appear OK > > You ran out of queue pairs. There is no way around this for larger all-to-all > transfers when using the openib btl and SRQ. Need O(cores^2) QPs to fully > connect with SRQ or PP QPs. I recommend using XRC instead by adding: > > btl_openib_receive_queues = X,4096,1024:X,12288,512:X,65536,512 > > > to your openmpi-mca-params.conf > > or > > -mca btl_openib_receive_queues X,4096,1024:X,12288,512:X,65536,512 > > > to the mpirun command line. > > > -Nathan > > On Jun 15, 2016, at 12:35 PM, "Sasso, John (GE Power, Non-GE)" > wrote: > > Chuck, > > The per-process limits appear fine, including those for the resource mgr > daemons: > > Limit Soft Limit Hard Limit Units > Max address space unlimited unlimited bytes Max core file size 0 0 > bytes Max cpu time unlimited unlimited seconds Max data size unlimited > unlimited bytes Max file locks unlimited unlimited locks Max file size > unlimited unlimited bytes Max locked memory unlimited unlimited bytes > Max msgqueue size 819200 819200 bytes Max nice priority 0 0 Max open > files 16384 16384 files Max pending signals 515625 515625 signals Max > processes 515625 515625 processes Max realtime priority 0 0 Max > realtime timeout unlimited
Re: [OMPI users] "failed to create queue pair" problem, but settings appear OK
XRC support is greatly improved in 1.10.x and 2.0.0. Would be interesting to see if a newer version fixed the shutdown hang. When calculating the required number of queue pairs you also have to divide by the number of queue pairs in the btl_openib_receive_queues parameter. Additionally Open MPI uses 1 qp/rank for connections (1.7+) and there are some in use by IPoIB and other services. -Nathan > On Jun 16, 2016, at 7:15 AM, Sasso, John (GE Power, Non-GE) >wrote: > > Nathan, > > Thank you for the suggestion. I tried your btl_openib_receive_queues > setting with a 4200+ core IMB job, and the job ran (great!). The shutdown > of the job took such a long time that after 6 minutes, I had to > force-terminate the job. > > When I tried using this scheme before, with the following recommended by the > OpenMPI FAQ, I got some odd errors: > > --mca btl openib,sm,self --mca btl_openib_receive_queues > X,128,256,192,128:X,2048,256,128,32:X,12288,256,128,32:X,65536,256,128,32 > > However, when I tried: > > --mca btl openib,sm,self --mca btl_openib_receive_queues > X,4096,1024:X,12288,512:X,65536,512 > > I got success with my aforementioned job. > > I am going to do more testing, with the goal of getting a 5000 core job to > run successfully. If I can, then down the road my concern is the impact the > btl_openib_receive_queues mca parameter (above) will have on lower-scale (< > 1024 cores) jobs, since the parameter change to the global openmpi config > file would impact ALL users and jobs of all scales. > > Chuck – as I noted in my first email, log_num_mtt was set fine, so that is > not the issue here. > > Finally, regarding running out of QPs, I examined the output of ‘ibv_devinfo > –v’ on our compute nodes. I see the following pertinent settings: > > max_qp: 392632 > max_qp_wr: 16351 > max_qp_rd_atom: 16 > max_qp_init_rd_atom:128 > max_cq: 65408 >max_cqe:4194303 > > Figuring that max_qp is the prime limitation here I am running into when > using the PP and SRQ QPs, considering 12 cores per node, this would seem to > imply that an upper bound on the number of nodes would be 392632 / 24^2 ~ 681 > nodes. This does not make sense, because I saw the QP creation failure error > (again, NO error about failure to register enough memory) for as small as 177 > 24-core nodes! I don’t know how to make sense of this, tho I don’t question > that we were running out of QPs. > > --john > > > From: users [mailto:users-boun...@open-mpi.org] On Behalf Of Nathan Hjelm > Sent: Wednesday, June 15, 2016 2:43 PM > To: Open MPI Users > Subject: EXT: Re: [OMPI users] "failed to create queue pair" problem, but > settings appear OK > > You ran out of queue pairs. There is no way around this for larger all-to-all > transfers when using the openib btl and SRQ. Need O(cores^2) QPs to fully > connect with SRQ or PP QPs. I recommend using XRC instead by adding: > > btl_openib_receive_queues = X,4096,1024:X,12288,512:X,65536,512 > > > to your openmpi-mca-params.conf > > or > > -mca btl_openib_receive_queues X,4096,1024:X,12288,512:X,65536,512 > > > to the mpirun command line. > > > -Nathan > > On Jun 15, 2016, at 12:35 PM, "Sasso, John (GE Power, Non-GE)" > wrote: > > Chuck, > > The per-process limits appear fine, including those for the resource mgr > daemons: > > Limit Soft Limit Hard Limit Units > Max address space unlimited unlimited bytes > Max core file size 0 0 bytes > Max cpu time unlimited unlimited seconds > Max data size unlimited unlimited bytes > Max file locks unlimited unlimited locks > Max file size unlimited unlimited bytes > Max locked memory unlimited unlimited bytes > Max msgqueue size 819200 819200 bytes > Max nice priority 0 0 > Max open files 16384 16384 files > Max pending signals 515625 515625 signals > Max processes 515625 515625 processes > Max realtime priority 0 0 > Max realtime timeout unlimited unlimited us > Max resident set unlimited unlimited bytes > Max stack size 30720 unlimited bytes > > > > As for the FAQ re registered memory, checking our OpenMPI settings with > ompi_info, we have: > > mpool_rdma_rcache_size_limit = 0 ==> Open MPI will register as much user > memory as necessary > btl_openib_free_list_max = -1 ==> Open MPI will try to allocate as many > registered buffers as it needs > btl_openib_eager_rdma_num = 16 > btl_openib_max_eager_rdma = 16 > btl_openib_eager_limit = 12288 > > > Other suggestions welcome. Hitting a brick wall here. Thanks! > > --john > > > > -Original Message- > From: users [mailto:users-boun...@open-mpi.org] On Behalf Of Gus Correa > Sent: Wednesday, June 15, 2016 1:39 PM > To: Open MPI Users > Subject: EXT: Re: [OMPI users] "failed to create queue pair" problem, but > settings appear OK
[OMPI users] "failed to create queue pair" problem, but settings appear OK
Nathan, Thank you for the suggestion. I tried your btl_openib_receive_queues setting with a 4200+ core IMB job, and the job ran (great!). The shutdown of the job took such a long time that after 6 minutes, I had to force-terminate the job. When I tried using this scheme before, with the following recommended by the OpenMPI FAQ, I got some odd errors: --mca btl openib,sm,self --mca btl_openib_receive_queues X,128,256,192,128:X,2048,256,128,32:X,12288,256,128,32:X,65536,256,128,32 However, when I tried: --mca btl openib,sm,self --mca btl_openib_receive_queues X,4096,1024:X,12288,512:X,65536,512 I got success with my aforementioned job. I am going to do more testing, with the goal of getting a 5000 core job to run successfully. If I can, then down the road my concern is the impact the btl_openib_receive_queues mca parameter (above) will have on lower-scale (< 1024 cores) jobs, since the parameter change to the global openmpi config file would impact ALL users and jobs of all scales. Chuck - as I noted in my first email, log_num_mtt was set fine, so that is not the issue here. Finally, regarding running out of QPs, I examined the output of 'ibv_devinfo -v' on our compute nodes. I see the following pertinent settings: max_qp: 392632 max_qp_wr: 16351 max_qp_rd_atom: 16 max_qp_init_rd_atom:128 max_cq: 65408 max_cqe:4194303 Figuring that max_qp is the prime limitation here I am running into when using the PP and SRQ QPs, considering 12 cores per node, this would seem to imply that an upper bound on the number of nodes would be 392632 / 24^2 ~ 681 nodes. This does not make sense, because I saw the QP creation failure error (again, NO error about failure to register enough memory) for as small as 177 24-core nodes! I don't know how to make sense of this, tho I don't question that we were running out of QPs. --john From: users [mailto:users-boun...@open-mpi.org] On Behalf Of Nathan Hjelm Sent: Wednesday, June 15, 2016 2:43 PM To: Open MPI Users Subject: EXT: Re: [OMPI users] "failed to create queue pair" problem, but settings appear OK You ran out of queue pairs. There is no way around this for larger all-to-all transfers when using the openib btl and SRQ. Need O(cores^2) QPs to fully connect with SRQ or PP QPs. I recommend using XRC instead by adding: btl_openib_receive_queues = X,4096,1024:X,12288,512:X,65536,512 to your openmpi-mca-params.conf or -mca btl_openib_receive_queues X,4096,1024:X,12288,512:X,65536,512 to the mpirun command line. -Nathan On Jun 15, 2016, at 12:35 PM, "Sasso, John (GE Power, Non-GE)"> wrote: Chuck, The per-process limits appear fine, including those for the resource mgr daemons: Limit Soft Limit Hard Limit Units Max address space unlimited unlimited bytes Max core file size 0 0 bytes Max cpu time unlimited unlimited seconds Max data size unlimited unlimited bytes Max file locks unlimited unlimited locks Max file size unlimited unlimited bytes Max locked memory unlimited unlimited bytes Max msgqueue size 819200 819200 bytes Max nice priority 0 0 Max open files 16384 16384 files Max pending signals 515625 515625 signals Max processes 515625 515625 processes Max realtime priority 0 0 Max realtime timeout unlimited unlimited us Max resident set unlimited unlimited bytes Max stack size 30720 unlimited bytes As for the FAQ re registered memory, checking our OpenMPI settings with ompi_info, we have: mpool_rdma_rcache_size_limit = 0 ==> Open MPI will register as much user memory as necessary btl_openib_free_list_max = -1 ==> Open MPI will try to allocate as many registered buffers as it needs btl_openib_eager_rdma_num = 16 btl_openib_max_eager_rdma = 16 btl_openib_eager_limit = 12288 Other suggestions welcome. Hitting a brick wall here. Thanks! --john -Original Message- From: users [mailto:users-boun...@open-mpi.org] On Behalf Of Gus Correa Sent: Wednesday, June 15, 2016 1:39 PM To: Open MPI Users Subject: EXT: Re: [OMPI users] "failed to create queue pair" problem, but settings appear OK Hi John 1) For diagnostic, you could check the actual "per process" limits on the nodes while that big job is running: cat /proc/$PID/limits 2) If you're using a resource manager to launch the job, the resource manager daemon/deamons (local to the nodes) may have to to set the memlock and other limits, so that the Open MPI processes inherit them. I use Torque, so I put these lines in the pbs_mom (Torque local daemon) initialization script: # pbs_mom system limits # max file descriptors ulimit -n 32768 # locked memory ulimit -l unlimited # stacksize ulimit -s unlimited 3) See also this FAQ related to registered memory. I set these parameters in /etc/modprobe.d/mlx4_core.conf, but where they're set may depend