We're excited to announce the release of AWS ParallelCluster 3.16.3
Upgrade
How to upgrade?
sudo pip install --upgrade aws-parallelcluster
CHANGES
- Upgrade EFA installer to 1.51.0 (from 1.49.0).
- Efa-driver: efa-3.3.1
- Efa-config: efa-config-1.18-1
- Efa-profile: efa-profile-1.7-1
- Libfabric-aws: libfabric-aws-2.7.0-1
- Rdma-core: rdma-core-65.0-1
- Open MPI: openmpi40-aws-4.1.8-1 and openmpi50-aws-5.0.9-11
- Upgrade NVIDIA driver, Fabric Manager, and IMEX to version 595.112.07 (from 595.71.05).
- Upgrade mysql-community-client to version 8.4.13 (from 8.4.10).
- Upgrade the Python version used by the virtual environments in the AMIs to 3.14.10 (from 3.14.8).
- Upgrade amazon-efs-utils to version 3.3.5 (from 3.1.3).
- Upgrade stunnel to version 5.82 (from 5.78).
- Upgrade Cinc Client to version 19.4.46 (from 19.3.14).
- Upgrade Amazon DCV to version 2026.0-25174 (from 2025.0-20103).
- server:
2026.0.25174-1 - xdcv:
2026.0.715-1 - gl:
2026.0.1124-1 - web_viewer:
2026.0.25174-1
- server:
- On Ubuntu, stop setting the kernel parameter
fs.protected_regularto 0, so that it keeps the OS default value as on the other OSes. - Grant
logs:DeleteLogGroupon the Log Group ARN instead of the Log Stream ARN to PClusterBuildImageCleanupRole
to prevent DELETE_FAILED of the build image stack after the upcoming CloudWatch Logs IAM change.
BUG FIXES
- Only enable NVIDIA IMEX service on GB200 compute nodes. Only enable NVIDIA Fabric Manager service on compute nodes with NVSwitch, and do not enable the
nvidia-powerdservice. Since 3.16.0 the services were enabled unnecessarily on all instance types. - Fix dynamic nodes getting stuck in
COMPLETINGstate indefinitely, holding their jobs, when the backing instance
of a node with jobs is terminated and its private IP is reused by another compute node. - Ensure only the latest AMI is returned by
pcluster list-official-imagesfor a given OS and architecture. - Fix name collisions of the
APIGatewayExecutionRoleIAM role when multiple ParallelCluster API stacks are created
at the same time, by using a longer suffix derived from the stack ID (24 characters, from 8).