Simone, If you want to run a single MPI task, you can either - mpirun -np 1 ./a.out (this is the most standard option) - ./a.out (this is the singleton mode. Note a.out will fork&exec an orted daemon under the hood, this is necessary for example if your app will MPI_Comm_spawn(). - OMPI_MCA_ess_singleton_isolated=1 ./a.out This is the singleton mode, and no orted daemon is spawned. this is faster but the app will abort if it invokes for example MPI_Comm_spawn().
mpirun does set an affinity mask by default. The default mask is one core per task is you run 2 MPI tasks or less, and a NUMA domain otherwise. Since you run a single MPI task and the default affinity mask is not a good fit, you can either run in singleton mode (and consider isolated singleton if it is a fit), or mpirun --bind-to none ... -np 1. Running one OpenMP thread per core vs hyperthread has to do with the OpenMP runtime and not Open MPI. Cheers, Gilles On Sun, Feb 17, 2019 at 12:15 PM Simone Atzeni <satz...@nvidia.com> wrote: > > Hi, > > > > For testing purposes I run some MPI+OpenMP benchmarks with `mpirun -np 1 > ./a.out`, and I am using OpenMPI 3.1.3. > > As far as I understand, `mpirun` sets an affinity mask, and the OpenMP > runtime (in my case the LLVM OpenMP RT) respects this mask and only sees 1 > physical core. > > In my case, I am running in a POWER8 which has 8 logical cores per physical > core. The OpenMP runtime in this case creates always the max number of > available logical cores in the machines (160 in my machine), but because of > the mask in this case will create 8 threads. > > All this threads runs in the same physical cores making the program slower > than if it would run each thread in a different physical core. > > > > So my question is, what’s the right way to run a single MPI process such that > the OpenMP threads can run in different physical cores independently from the > mask set by mpirun? > > > > I know about the option `--bind-to none` and using that all the cores in the > system become available and the OpenMP runtime uses all of them. > > Otherwise, doing some web search I read that a singleton MPI program should > be executed with ` OMPI_MCA_ess_singleton_isolated=1 ./a.out` without > `mpirun` at all, but I couldn’t find a good explanation of it. > > > > Is there anyone that could clarify this? > > > > Thank you! > > Simone > > > > ________________________________ > This email message is for the sole use of the intended recipient(s) and may > contain confidential information. Any unauthorized review, use, disclosure > or distribution is prohibited. If you are not the intended recipient, please > contact the sender by reply email and destroy all copies of the original > message. > ________________________________ > _______________________________________________ > users mailing list > users@lists.open-mpi.org > https://lists.open-mpi.org/mailman/listinfo/users _______________________________________________ users mailing list users@lists.open-mpi.org https://lists.open-mpi.org/mailman/listinfo/users
