Check whether NVMe health polling is disturbing an L-series storage test

Source facts

Microsoft warns against mixing NVMe administrative commands, including SMART queries, with active I/O on Lsv3, Lasv3, and Lsv2 VMs. Their NVMe Direct implementation switches to a slower mode while administration commands are pending. The documented maximum queue depth is 1,024 per I/O queue pair; exceeding it in a synthetic benchmark can cause queue-full conditions and reduce performance. Local NVMe storage is ephemeral. Microsoft Learn.

Applicability

Review the benchmark harness and monitoring agents together when a Windows workload on these families behaves differently under observation. Distinguish an application change from a change in the measurements and background queries surrounding it.

DSE recommendation

Separate health-query activity from the workload measurement before diagnosing the storage limit. Ask the performance owner to identify every source of NVMe administrative polling during the test. Schedule a controlled comparison that changes the relevant observation activity while preserving workload parameters. Keep the test’s queue settings within the documented limit and use only disposable or appropriately protected data.

Verification

Record the workload, queue configuration, administrative-query timing, and measured results for each run. Look for a repeatable relationship before assigning a cause; do not claim that all monitoring tools or all low performance have this mechanism. Restore the approved monitoring configuration after the test, and retain the evidence for any proposed monitoring change separately from the storage-capacity plan.

Official references

Microsoft Learn: Optimize performance on Lsv3, Lasv3, and Lsv2 Windows VMs. Source reviewed September 9, 2026.