Measure DOCA library throughput, latency, and bandwidth with the doca_bench CLI harness.
npx skills add https://github.com/NVIDIA/skills --skill doca-bench
Please help me install this Agent Skill. Skill: doca-bench Source: https://github.com/NVIDIA/skills/tree/main/skills/doca-bench Command: npx skills add https://github.com/NVIDIA/skills --skill doca-bench
Measuring the real performance of NVIDIA DOCA libraries (RDMA, Compress, AES-GCM, SHA, DMA, EC, Ethernet, Comch, GPUNetIO) on a specific host or BlueField Arm device is error-prone: agents and operators guess numbers from datasheets, invent non-existent CLI flags, skip warm-up, and quote results without the version/device context needed to reproduce them. This Skill guides an agent through the shipped doca_bench micro-benchmark harness so measurements are configured, executed, and reported correctly.
Ask your agent to measure the throughput of DOCA Compress on your BlueField device using doca_bench and capture a reproducible baseline.
High-intent search queries and answers about installing and using this skill.
Use doca_bench by committing to three axes: the target DOCA library, the workload shape (direction, data provider, batching, NUMA placement), and the measurement mode (throughput, bulk-latency, precision-latency, or max-bandwidth). Always run a short smoke test before any long or swept run.
doca_bench ships four benchmark modes: throughput (aggregate op-rate), bulk-latency (batched latency distribution), precision-latency (per-job percentiles), and max-bandwidth (saturation ceiling). The modes are not interchangeable, so numbers from different modes must not be compared directly.
doca_bench is available since DOCA 2.7.0 and ships at /opt/mellanox/doca/tools/doca_bench on every install, host or BlueField Arm. On older installs the binary is absent, so the fix is upgrading DOCA rather than using alternative tools.
Zero or hung runs usually come from device-binding or library-precondition failures, not measurement issues. Confirm the device is visible to DOCA with doca_caps, then re-run the bench query system to verify the granular build actually exposes the target library.
No. doca_bench measures the DOCA library surface only, not application-level end-to-end timing. For application measurements, write a custom program against the relevant DOCA library using the doca-programming-guide and per-library skills.
Capture the CSV output plus the four-tuple: exact command line, DOCA version, device target, and as-deployed environment (firmware, kernel, NUMA, hugepages). Require two consecutive runs within a predeclared tolerance before calling the baseline stable.
air-gappedAutomate NVIDIA datacenter GPU bring-up on Ubuntu 24.04 with health validation.
r3bl-orgGenerate and compare flamegraphs against baselines to detect performance regressions.
nguyenpv1980-wqReview skill-library pull requests for registration consistency and collision safety.
zhuangbiaoweiEvaluate code generation models on HumanEval, MBPP, and MultiPL-E using pass@k metrics.
