RELEASED: The HPE Supercomputing Programming Environment (SUPS) 1.0.0 General Availability Release for HPE Supercomputing EX systems with RHEL 9.6 – October 2026

PRODUCT OVERVIEW

HPE Supercomputing Programming Software (SUPS) provides a validated and supported programming environment for developing, optimizing, and deploying HPC and AI applications. SUPS combines compilers, libraries, communication technologies, debugging and performance tools into reproducible software environments designed to reduce integration complexity, accelerate development, and simplify software lifecycle management across modern CPU and GPU architectures.

Documentation

NOTE: SUPS 1.0.0 (and later) release announcement and user guides, along with installation guides, are now available on the CPE Documentation site. CPE installation guides for all CPE versions can still be accessed from the HPE Support Center web page.

IMPORTANT NOTES

CMA Support with hpc-container-wrapper Limitation

When HPE Cray MPICH applications are launched using the hpc-container-wrapper script, Linux Cross Memory Attach (CMA) might not be available for intra-node MPI communication.

The hpc-container-wrapper creates a separate user namespace for each MPI rank. CMA relies on Linux process_vm_readv() and related cross-process memory access mechanisms. When ranks execute in different user namespaces, the operating system might deny these operations and cause CMA initialization to fail with errors, such as:

PE 0: CMA does not have sufficient permission to operate:
Consider disabling CMA by setting MPICH_SMP_SINGLE_COPY_MODE=NONE as a workaround.
process_vm_readv: Operation not permitted

As a result, applications that normally use CMA for intra-node transfers might abort during MPI initialization or communication setup.

On most systems, XPMEM is the default and preferred intra-node transport with HPE Cray MPICH. CMA is used only when XPMEM is unavailable or has been disabled.

Shared Memory Implementation Workaround

Disable the HPE Cray MPICH single-copy transport selection, and force use of the shared memory implementation:

export MPICH_SMP_SINGLE_COPY_MODE=NONE

This implementation causes HPE Cray MPICH to bypass both XPMEM and CMA and use the two-copy shared-memory implementation instead. This workaround, however, might negatively impact intra-node MPI performance, particularly for large message transfers. The shared memory implementation requires additional memory copies compared to the CMA and XPMEM single-copy transports.

Automated CMA Selection on Select GPU Systems

When running GPU-enabled applications on systems with AArch64 CPUs and NVIDIA GPUs, HPE Cray MPICH automatically disables XPMEM and selects CMA for certain intra-node host-memory transfers. In these environments, the hpc-container-wrapper limitation can be more significant. If CMA cannot be used due to container namespace restrictions, the only remaining option is the shared-memory transport, which might result in reduced intra-node communication performance.

SUPS 1.0.0 Product Versions:

The HPE Supercomputing Programming Environment (SUPS) for HPE Supercomputing EX Systems with HPCM

  • cce-22.0.0

  • cray-mpich-9.2.0

  • cray-pmi-6.1.18

  • sups-1.0.0