Personal collection of HPC communication benchmarks covering SHMEM, MPI, UCX, UCC, and SHARP.
Actively maintained suites are marked with active; others are preserved for historical reference. This repository informs upstream discussions — see openucx/ucc PRs for related work.
| Dir | What it measures | Build requirements |
|---|---|---|
ucc/ active |
UCC collectives: alltoall, alltoallv (fixed + variable), pipelined a2av bandwidth, context/team creation timing | MPI, UCC, UCX |
ucx/ |
UCX-level alltoall and full HPCAC exchange patterns | UCX, MPI |
mpi/ |
RDMA all-to-all variants, MPI alltoall benchmark, scalability analysis | MPI (RDMA-capable) |
shmem/ |
SHMEM put/stencil/nonblocking/cuda benchmarks; Graph500; comprehensive test suites | SHMEM (SYCL/MPI-based) |
sharp/ |
Mellanox SHARP locality benchmark for in-network reduction | SHARP, MPI |
stream/ |
Endless copy/stream throughput micro-benchmark | none |
archive/ |
Historical results (e.g. 2019 SHMEM alltoall experiments) | — |
# UCC alltoall
cd ucc/a2a && make && mpirun -np 4 ./ucc_bench_a2a
# UCC context/team timing
cd ucc/timing-ucc && make && mpirun -np 4 ./time-ucc
# UCX-level benchmarks
cd ucx/alltoall && make && mpirun -np 4 ./alltoall
# MPI RDMA alltoall
cd mpi && make && mpirun -np 4 ./rdma_alltoall_demoAll UCC benchmarks emit CSV — see docs/output-format.md.
Historical results under archive/.