ChecklistInformationITNetworks & Infrastructure

Check the InfiniBand placement boundary before troubleshooting MPI reachability

Are the HPC VMs positioned to use Azure's RDMA fabric rather than only sharing ordinary Ethernet connectivity?

Resilient network core with engineered blue and gold data paths.
DSE visual intelligenceNetworks & infrastructureChecklist · 2 min read
Executive summary

What you need to know

Are the HPC VMs positioned to use Azure's RDMA fabric rather than only sharing ordinary Ethernet connectivity?

Potentially affected

Owners of RDMA-capable Azure HPC deployments checking InfiniBand placement and network addressing.

DSE recommendation

Map each communicating node to its actual scale set or availability set before changing MPI settings.

Source facts

Azure’s documented HPC RDMA path uses InfiniBand, not the ordinary Ethernet interface. VMs needing RDMA communication must share a scale set or availability set; a scale-set deployment requires one placement group for its InfiniBand communication. Microsoft reserves 172.16.0.0/16 for the RDMA network and warns against overlapping the virtual network with it. Appropriate InfiniBand drivers are also required. Microsoft Learn.

Applicability

Confirm the exact RDMA-capable VM size, supported operating system, driver stack, and deployment model. The source contains older platform examples; use its current linked size and image guidance when qualifying a new build. Do not infer fabric eligibility solely from successful ordinary network tests.

DSE recommendation

Map each communicating node to its actual scale set or availability set before changing MPI settings. Have the compute owner confirm the scale-set placement configuration and the network owner compare address ranges with the reserved fabric space. Resolve a placement or address-design mismatch through an approved deployment plan, preserving the application’s existing recovery path.

Verification

On designated test nodes, inspect the actual fabric interface and driver state, then exercise the intended multi-node workload. Capture placement identifiers, selected interface, and observed communication separately from an Ethernet reachability result. If the job still fails within a valid fabric boundary, pass that evidence to the MPI owner for library-specific investigation. Avoid changing several placement, driver, and runtime variables together without a way to attribute the result.

Official references

Microsoft Learn: Set up InfiniBand. Source reviewed September 9, 2026.

Primary reference

Review the official source

Set up InfiniBand on HPC VMs - Azure Virtual Machines - Azure Virtual Machines | Microsoft Learn · Verified September 9, 2026

Open official reference ↗
Plan the next step

Need help applying this guidance safely?

DSE can help confirm applicability, protect service continuity, and validate the result across physical security and IT systems.

Talk with DSE