Changes and New Features

HPC-X Content

Updated HPC-X Content section to reflect the communication libraries versions embedded in this HPC-X release:

• IMB Benchmarks v2021.10

• nccl_spcx_plugin v1.4.0

• NVIDIA SHARP v3.15

• UCX v1.21

• UCC v1.9.0

UCC: TL/UCP Strided Active Set Broadcast

Added support for broadcast collectives over strided active sets larger than two ranks. This allows for flexible rank subsetting using start, stride, and size parameters. For further information, please see TL/UCP Strided Active Set.

UCC

Starting with HPC-X v2.50, UCC is the default collective framework for Open MPI collectives. HCOLL is included in HPC-X but is disabled by default.

Improved UCC correctness on NVIDIA Grace Hopper systems by strengthening AArch64 memory ordering and shared-memory collective synchronization.

Open MPI

Open MPI v5.0.x is now the default MPI implementation in HPC-X. Open MPI v4.1.x is available and can be set using an environment variable - see Installing and Loading HPC-X for further details.

Benchmarks

IMB Benchmarks: Upgraded to v2021.10.

• Improved GPU startup by initializing CUDA prior to MPI_Init. This aligns with standard CUDA-aware MPI application behavior and prevents late-initialization errors.