Naive benchmark of a toy MPI application in which we place the Julia depot (i.e. JULIA_DEPOT_PATH) on different file systems at the Paderborn Center for Parallel Computing (PC2). Run on the normal compute nodes of Noctua 2 (dual-socket AMD EPYC Milan 7763 64-Core CPUs).
Heavily inspired by a similar Python benchmark run at NERSC: https://gitlab.com/NERSC/python-benchmark/-/tree/main/shifter-vs-filesystems.
Module: mpi/OpenMPI/4.1.1-GCC-11.2.0
| Number of nodes | PC2PFS | HOME | Container |
|---|---|---|---|
| 1 | 0m5.255s | 0m5.713s | tbd |
| 2 | 0m5.411s | 0m6.149s | tbd |
| 10 | 0m5.705s | 0m7.529s | tbd |
| 100 | 0m11.984s | 0m18.926s | tbd |
| 500 | 0m34.236s | 2m36.020s | tbd |
Module: mpi/impi/2021.5.0-intel-compilers-2022.0.1
| Number of nodes | PC2PFS | HOME | Container |
|---|---|---|---|
| 1 | 0m5.077s | 0m5.659s | tbd |
| 2 | 0m5.120s | 0m5.961s | tbd |
| 10 | 0m6.927s | 0m8.362s | tbd |
| 100 | 0m15.810s | 0m22.069s | tbd |
| 500 | 0m23.908s | 2m18.505s | tbd |