selftests/resctrl: Run the MBA test on arm64 MPAM via the NVIDIA SCF PMU
Get series with b4
b4 am [email protected]Versions newest first
- v1 Latest5 patches 2 repliesIn review
Patches in v1 No applies recorded yet
| # | Subject and files | Size | Replies | Applied |
|---|---|---|---|---|
| 1/5 | selftests/resctrl: Make memory bandwidth measurement vendor-neutraltools/ tools/ tools/ +1 more | +114 | 0 replies | |
| 2/5 | selftests/resctrl: Abstract reference bandwidth counters behind a backendtools/ tools/ tools/ +1 more | +141 | 0 replies | |
| 3/5 | selftests/resctrl: Support the MPAM memory bandwidth monitoring layouttools/ tools/ tools/ | +111 | 0 replies | |
| 4/5 | selftests/resctrl: Gate the MBA test on features, not CPU vendortools/ | +6 | 0 replies | |
| 5/5 | selftests/resctrl: Add NVIDIA SCF reference bandwidth backendtools/ tools/ | +162 | 0 replies |
Reviews trailers per patch
No Reviewed-by, Acked-by or Tested-by trailers for v1 yet.
Cover letter [PATCH 0/5] selftests/resctrl: Run the MBA test on arm64 MPAM via the NVIDIA SCF PMU
The resctrl MBA selftest is currently Intel-only: the test is vendor-gated to GenuineIntel and its reference bandwidth measurement is hard-wired to the Intel iMC uncore PMU. MPAM on arm64 exposes the same MB resource through resctrl, and NVIDIA Grace/Vera systems provide an uncore PMU (the SoC coherency fabric, SCF) that can serve as the same kind of independent reference that the iMC provides on x86.
Patches 1-2 decouple the test machinery from the iMC with no functional change and no NVIDIA code: a mechanical rename, then a small backend abstraction (detect()/setup_counters()) with the iMC as the only backend. Patch 3 teaches the framework the MPAM memory bandwidth monitoring layout (NUMA-node-keyed MB domains, MB_MON's mbm_total_bytes, ABMC counter assignment). Patch 4 switches the MBA test from vendor gating to feature detection. Patch 5 adds the NVIDIA SCF PMU as a reference bandwidth backend, making the MBA test runnable on arm64.
The series builds after each patch, and mid-series behavior is sane:
on arm64 the MBA test skips cleanly until patch 5, after which it runs and passes. x86 behavior is preserved throughout - the iMC stays the default backend and the mbm_local_bytes comparison path is unchanged.
This series applies on top of the 26.04_linux-nvidia.glue.others.cpu_less.hardlimit branch of
https://github.com/fyu1/NV-Kernels.fenghuay.baseos
The base-commit recorded below is the current tip of that branch.
Validated on a two-socket VR-NVL72 (Vera) system: the SCF-measured bandwidth and resctrl mbm_total_bytes agree within 3-4% across all MBA schemata levels (10%..100%), against the test's 8% threshold.
Richard Cheng (5):
selftests/resctrl: Make memory bandwidth measurement vendor-neutral selftests/resctrl: Abstract reference bandwidth counters behind a
backend
selftests/resctrl: Support the MPAM memory bandwidth monitoring layout selftests/resctrl: Gate the MBA test on features, not CPU vendor selftests/resctrl: Add NVIDIA SCF reference bandwidth backend
tools/testing/selftests/resctrl/mba_test.c | 38 +- tools/testing/selftests/resctrl/mbm_test.c | 23 +- tools/testing/selftests/resctrl/resctrl.h | 6 +- tools/testing/selftests/resctrl/resctrl_val.c | 530 ++++++++++++++---- tools/testing/selftests/resctrl/resctrlfs.c | 45 +- 5 files changed, 519 insertions(+), 123 deletions(-)
base-commit: 475f355d0408698d9f632ba3b9a5dfb1f91e928b
Signature · 1 line
2.43.0