ParallelCluster 3.16 drops Amazon Linux 2 and AWS Batch, 20 Aug 2026

Summary. AWS ParallelCluster 3.16.0 was tagged on 20 August 2026 and announced on 24 August 2026 under the headline "adds an on-node diagnostics tool." The announcement mentions pcluster-diag, "cluster stability improvements, and an updated HPC and AI/ML software stack." It does not mention that the same release ends support for Amazon Linux 2, ends support for AWS Batch as a scheduler, changes the default NFS lock manager port from 32768 to 4045, and adds a required IAM permission. All four appear in the AWS ParallelCluster User Guide release notes and in the project CHANGELOG on GitHub. If your cluster runs Os: alinux2 or Scheduler: awsbatch, 3.16.0 is not an upgrade, it is a rebuild. This is the sort of release where the changelog is the announcement and the announcement is marketing.

What AWS said, and what the release notes say

The AWS What's New post, dated 24 August 2026, describes 3.16 in five sentences. The ParallelCluster User Guide release notes date the same version 20 August 2026 and run to roughly sixty line items, ending in a two-line Deprecations block. The GitHub release for v3.16.0 was published on 20 August 2026 and carries the same list.

The diagnostics tool itself is real and useful: AWS documents pcluster-diag as context aware, read-only by default, and included in every ParallelCluster AMI from 3.16.0 onward, official and custom, with a describe-checks subcommand that lists every registered check as JSON. The gap is not a documentation bug. The What's New feed is a launch channel and the release notes are the contract. Read the contract.

The two deprecations that end a cluster

Amazon Linux 2 is gone as a supported operating system. Any cluster configuration carrying Os: alinux2 has to move to another supported operating system before the head node image is rebuilt. The operating system is set in the Image section of the cluster configuration file, and the same user guide notes that unsupported official AMIs "will be made unavailable after 18 months of inactivity." This is a create-new-cluster migration with data movement, not a pcluster update-cluster.

AWS Batch as a scheduler is also gone. The Scheduling section of the user guide still documents AwsBatchQueues alongside SlurmQueues, and now carries the line "Starting with AWS ParallelCluster version 3.16.0, AWS Batch as a scheduler is no longer supported." The same page settles the upgrade question: the update policy for the scheduler setting reads "If this setting is changed, the update is not allowed." You cannot switch a Batch cluster to Slurm with pcluster update-cluster. The two deprecations also interact, because the docs note that "awsbatch only supports the alinux2 operating system and x86_64 platform," so a Batch cluster fails both checks at once. Slurm in 3.16.0 is version 25.11.6, up from 25.11.4.

The NFS port change that will look like a network fault

This is the one that generates a support ticket rather than a build failure. From the release notes: the default NFS lock manager (lockd) port moves from 32768 to 4045 "to avoid collisions with the Linux ephemeral port range (32768-60999), which can cause sporadic mount failures." AWS then states the consequence plainly: "If you restrict NFS ports in a firewall, you must open TCP/UDP 4045 or NFSv3 mounts will fail."

Two details decide whether this hits you. First, ParallelCluster-managed storage is mounted over NFSv4 and is unaffected, because 3.16.0 also enforces NFSv4-only on the managed NFS server running on the head node. Second, the NFSv3 client stack (rpcbind, rpc-statd, lockd) is unchanged, so nodes that mount an external NFSv3 server still use it. The failure mode is therefore narrow and nasty: an existing external NFSv3 filer, an existing security group that was correct yesterday, and mounts that fail after an image rebuild.

The amazon-efs-utils row is another quiet change: 3.16.0 installs it from the official Amazon EFS endpoint instead of building from source, and the release notes state that users "must now allowlist the CloudFront domain amazon-efs-utils.aws.com in their proxy/egress configuration when using build-image with isolated subnets."

Three more things that break before your job runs

The ClusterNameValidator now caps cluster names at 40 characters when ExternalSlurmdbd is configured, matching the existing limit for Database. AWS gives the reason: MySQL table name length. A 45-character cluster name that worked on 3.15.1 now fails validation.

The CLI requires the additional IAM permission tag:GetResources, used to resolve login node load balancer ARNs by tags. If your ParallelCluster IAM policy is a hand-maintained least-privilege document rather than the AWS-managed one, pcluster calls will fail until you add it.

ParallelCluster-managed bootstrap files move off /tmp into /opt/parallelcluster/tmp, so image builds, cluster creation and updates now work on custom AMIs that mount /tmp with noexec. The release notes attach a condition: "Image builds work on the custom AMIs only if GDRcopy installation is skipped." A hardened-AMI shop gets the fix and loses GDRcopy in the same sentence.

The security fix nobody labelled a security fix

Buried in Bug Fixes: "Prevent a path traversal when downloading exported cluster and image logs, so a CloudWatch log stream name containing ../ segments can no longer cause files to be written outside the export directory." That is a path traversal in a CLI that HPC administrators run with elevated local privileges, fixed without a CVE reference or a security bulletin in the What's New post. Anyone running pcluster export-cluster-logs against a cluster whose log stream names are influenced by job metadata should treat 3.16.0 as a security update, not a feature release.

The software stack, version by version

Enroot 3.4.1 to 4.2.1 is a major version jump inside a point release of ParallelCluster, and Pyxis moves with it. Container-based Slurm workloads should be re-tested rather than assumed. NVIDIA components now install from distribution package manager local repo packages instead of run file installers, and on RHEL-family systems that change forces the Xorg driver for DCV GPU acceleration with Wayland disabled so GDM always starts Xorg. The CLI also gains Python 3.13 support.

How to tell if this is you

1. grep -r 'alinux2' across your cluster configuration files. Any hit is a rebuild.
2. grep -r 'awsbatch' in the same place. Any hit is a scheduler migration.
3. Check whether any compute node mounts an external NFSv3 export. If yes, open TCP and UDP 4045 before rebuilding the AMI.
4. Diff your ParallelCluster IAM policy against the 3.16.0 requirement and add tag:GetResources.
5. Count the characters in your cluster name if you use ExternalSlurmdbd. Over 40 fails.

India-specific considerations

Indian research computing groups and engineering services teams running ParallelCluster in ap-south-1 tend to sit behind corporate proxies with strict egress allowlists, which makes the amazon-efs-utils.aws.com requirement and the 4045 port change the two most likely upgrade blockers here rather than the deprecations. Budget the Amazon Linux 2 migration as a project, not a maintenance window: the operating system change forces a new cluster, and shared filesystem data has to move with it. Teams already tracking AWS Savings Plans coverage sizing should re-check commitments before rebuilding, because a rebuild changes instance families in practice more often than plans assume, and our notes on shared GPU tenancy controls apply to the refreshed NVIDIA stack.

What is still unknown

AWS has not published an end-of-support date for ParallelCluster 3.15.x, so there is no stated window in which an Amazon Linux 2 or AWS Batch cluster can keep receiving fixes on the older line. The release notes also do not say whether the path traversal fix will be backported. The CHANGELOG on the project's develop branch already carries an unreleased 3.17.0 section listing export-logs improvements, which suggests the log export path is still being worked on.

FAQ

When was AWS ParallelCluster 3.16.0 released?

The GitHub release for v3.16.0 was published on 20 August 2026 and the AWS ParallelCluster User Guide release notes carry the same date. The AWS What's New announcement appeared four days later, on 24 August 2026, under a headline naming only the new on-node diagnostics tool.

Does ParallelCluster 3.16 still support Amazon Linux 2?

No. The release notes list "Amazon Linux 2 is no longer supported" under Deprecations for 3.16.0. Because the operating system is a cluster property in ParallelCluster rather than something you change in place, moving off it means creating a new cluster and migrating shared filesystem data across.

Can I still use AWS Batch as a ParallelCluster scheduler?

No. The 3.16.0 Deprecations block states that AWS Batch as a scheduler is no longer supported. Existing Batch-backed deployments have to move to Slurm, which in 3.16.0 is version 25.11.6. Job submission wrappers written against the Batch CLI commands need rewriting for Slurm.

Why did the NFS lockd port change to 4045?

AWS moved the default NFS lock manager port from 32768 to 4045 to avoid collisions with the Linux ephemeral port range of 32768 to 60999, which caused sporadic mount failures. The change affects only nodes mounting an external NFSv3 server, because ParallelCluster-managed storage now uses NFSv4 exclusively.

What new IAM permission does the 3.16 CLI need?

The CLI requires tag:GetResources, which it uses to resolve login node load balancer ARNs by tags. The same underlying change fixes a race condition in load balancer lookup that previously caused cluster update failures on clusters with login nodes. Hand-maintained least-privilege policies must add it.

Is there a security fix in ParallelCluster 3.16.0?

Yes, listed under Bug Fixes rather than announced separately. AWS prevented a path traversal when downloading exported cluster and image logs, so a CloudWatch log stream name containing dot-dot-slash segments can no longer write files outside the export directory. No CVE reference appears in the release notes or the announcement.

Which GPU and networking components changed in 3.16.0?

The NVIDIA driver, Fabric Manager and IMEX move to 595.71.05 from 580.126.20, CUDA Toolkit to 13.2.2 from 13.0.2, DCGM to 4.6.0 from 4.5.1, and the EFA installer to 1.49.0 from 1.47.0. NVIDIA components now install from distribution package manager local repo packages instead of run file installers.

What is the cluster name length limit in 3.16.0?

The ClusterNameValidator now limits cluster names to 40 characters when ExternalSlurmdbd is used, matching the limit that already applied when Database is configured. AWS gives the reason as MySQL table name length, and says the validator prevents runtime failures that previously surfaced only after the cluster was built.

How eCorpIT can help

eCorpIT runs AWS platform migrations for engineering and research teams, and our senior engineering teams read release notes against live configuration rather than against announcements. We are CMMI Level 5 and ISO 27001:2022 certified. If you have ParallelCluster stacks on Amazon Linux 2 or AWS Batch and need the migration scoped before support pressure arrives, book an AWS HPC migration review.

References

1. AWS ParallelCluster 3.16 adds an on-node diagnostics tool, AWS What's New, 24 August 2026
2. Release notes and document history, AWS ParallelCluster User Guide
3. aws-parallelcluster CHANGELOG, GitHub
4. aws-parallelcluster latest release metadata, GitHub API
5. Scheduling section, AWS ParallelCluster User Guide
6. Image section, AWS ParallelCluster User Guide
7. IAM permissions in AWS ParallelCluster
8. Troubleshooting with pcluster-diag, AWS ParallelCluster User Guide
9. SharedStorage section, AWS ParallelCluster User Guide
10. pcluster export-cluster-logs, AWS ParallelCluster User Guide
11. Cluster configuration file, AWS ParallelCluster User Guide
12. AMI customization considerations, AWS ParallelCluster User Guide
13. What is AWS ParallelCluster, AWS ParallelCluster User Guide

Last updated: 24 August 2026.