OSU Micro-Benchmarks
OSU Micro-Benchmarks calibrate MPI interconnect performance. On the Orange Pi RV2 we measure on-node OpenMPI shared-memory (self+vader) so MPI-heavy apps (HPL, ScaLAPACK, QE, PETSc) can be read against a known latency/bandwidth floor.
Benchmark source: opensolvers/benchmarks/osu. Module: OSU-Micro-Benchmarks/7.5.1-gompi-2025b.
Bottom line: sub-2 μs shared-memory latency and ~2 GB/s uni-directional BW — pure MPI overhead is small next to BLAS/FFT work on this board. Cross-node Ethernet needs a second machine (not measured here).
Results (2026-08-21)
OpenMPI 5.0.8, pml=ob1 btl=self,vader, --bind-to core.
Point-to-point (np=2)
| Metric | Value |
|---|---|
| Latency @ 1 B | 1.12 μs |
| Latency @ 1 KiB | 2.73 μs |
| Latency @ 1 MiB | 640.56 μs |
| Uni-directional BW peak | 2069 MB/s (@ 512 KiB) |
| Bi-directional BW peak | 2456 MB/s (@ 256 KiB) |
Collectives (np=8, full X60)
| Test | latency @ 4 B / 1 KiB / 1 MiB |
|---|---|
osu_allreduce |
6.96 / 18.51 / 13410 μs |
osu_bcast |
3.04 / 8.86 / 3758 μs |
osu_alltoall |
10.97 / 41.90 / 24411 μs |
When an 8-rank ScaLAPACK/PETSc run looks communication-bound, compare message sizes to the collective table — alltoall at ≥64 KiB is already multi-millisecond per call.
Measured: 2026-08-21 on Orange Pi RV2.