Discover the HPC fabric partition key instead of copying an MPI example value

Source facts

Microsoft notes that some MPI libraries on SR-IOV-enabled VMs require partition-key configuration. Its discovery procedure reads two fabric port key entries and selects the larger value as the tenant key. That value remains stable while the availability set or scale set exists, including node additions and removals; a new tenant receives different keys. The documented snippets are examples, not current-version mandates. Microsoft Learn.

Applicability

Use this review when the selected MPI library needs a partition key. Confirm the runtime, fabric interface name, and actual deployment boundary before adapting the discovery procedure. Keep this investigation separate from whether the nodes were placed together and have supported InfiniBand drivers.

DSE recommendation

Discover the deployed tenant’s partition key and record its source instead of retaining the sample constant. Ask the HPC owner to distinguish a node replacement within the same boundary from rebuilding the workload in a new one. Include rediscovery in the latter handoff and review the runtime’s specific way of accepting the value. Do not apply unrelated historical package or SSH settings merely because they share the example page.

Verification

Compare the configured runtime value with the discovered entries on approved test nodes. Execute a bounded communication test using the intended library, retaining runtime and deployment identifiers with the outcome. Recheck the value after a deployment-boundary replacement, while avoiding unnecessary changes for an ordinary node addition. Preserve enough evidence to separate an incorrect partition selection from placement, driver, or application failures.

Official references

Microsoft Learn: Set up MPI for HPC. Source reviewed September 9, 2026.