The performance we observe on NVL72 is still suboptimal, with NCCL being roughly 2x faster. We are currently at ~20GiB/s, where NCCL achieves ~40GiB/s on 64 GPUs.
We need to further investigate where we are losing. Most obvious probable candidates are:
- RapidsMPF
bench_comm
- UCXX lock contention
- UCX internals (could be a variety of different reasons)
Possible ways to explore this further:
- All-to-all test with OSU+UCX
- Properly configure to use MPI communicator with UCX backend on NVL72 (currently hitting issues with MLX5, probably due to minor misconfigurations that simply need to be corrected)
The performance we observe on NVL72 is still suboptimal, with NCCL being roughly 2x faster. We are currently at ~20GiB/s, where NCCL achieves ~40GiB/s on 64 GPUs.
We need to further investigate where we are losing. Most obvious probable candidates are:
bench_commPossible ways to explore this further: