Matias, Assuming you run one MPI task per unikernel, and two unikernels share nothing, it means that inter-node communication cannot be performed via shared memory or kernel feature (such as xpmem or knem). That also implies communication are likely using the loopback interface which is much slower.
Cheers, Gilles On Sun, Jul 24, 2022 at 8:11 PM Matias Ezequiel Vara Larsen via users < users@lists.open-mpi.org> wrote: > Hello everyone, > > I have started to play with MPI and unikernels and I have recently > implemented a minimal set of MPI APIs on top of Toro Unikernel > ( > https://github.com/torokernel/torokernel/blob/example-mpi/examples/MPI/MpiReduce.pas > ). > I was wondering if someone may be interested in the use of unikernels to > deploy MPI applications. Toro is a shared-nothing unikernel in which > each core runs independently from one another. Also, memory is per-core > to leverage NUMA. I was thinking that those features may improve the > execution of MPI applications but I have not measured that yet. For the > moment, I am running a simple MPI reduction with the MPI_SUM operation > and watching how this behaves when the number of cores increases. Do you > know any benchmark that I can run so try to test that? > > Matias >
