DOCA Flow Performance Measurement
Guides reproducible, defensible measurements of host-side and DPU-CPU DOCA Flow control-plane rule rates.
The documentation specifies scoped execution, user confirmation, dedicated-device checks, OOB access, smoke-before-bulk testing, evidence capture, and disclosure that real Flow rules consume device resources. The main safety risks are therefore visible. Seven points are deducted because rollback/cleanup procedures, dependency-security review, and systematic sensitive-data handling are incomplete, while large runs may still affect production hardware.
TASKS.md gives a structured configure, smoke, iterative-test, and layered-debug workflow with reasonably specific failure feedback. However, exact CLI flags and the JSON schema are delegated to external help/docs, no executable test suite or key-path reproduction evidence is supplied, and part of the provided content is abstract or truncated. Twelve points are deducted under the static-review ceiling.
Triggers, audience, use cases, and boundaries versus doca-flow-tune and doca-flow-dpa-perf are clearly stated. The capability nevertheless depends heavily on Linux, the DOCA SDK, BlueField/ConnectX hardware, and locally installed binaries; Chinese-language guidance is absent, and reachability of external NVIDIA documentation from mainland-China networks is not established. Five points are deducted.
SKILL.md, CAPABILITIES.md, and TASKS.md provide layered information architecture with dependencies, limitations, version matching, routing, and maintenance contacts. Six points are deducted for the absence of a clear author, changelog, and stable version policy, inconsistencies between SKILLCARD metadata and the supplied repository/license metadata, and reported missing recommended sections in static validation.
The skill directly helps with policy selection, backend choice, workload configuration, smoke testing, iterative measurement, and reproducibility capture, while correctly separating control-plane rate from dataplane throughput. Nine points are deducted because exact commands, flags, and schema details are not supplied, requiring external documentation and substantial environment-specific validation.
The files define error categories, raw-output and configuration capture requirements, and include a claimed passing evaluation report. Seven points are deducted because the evaluation dataset is unavailable, no real CI workflow and committed key-path test suite are provided, and the report is primarily an author-maintained summary that cannot be independently reproduced from the supplied material.
- This is a static documentation review; no doca_flow_perf command, test, or hardware operation was executed.
- The workflow may install and delete large numbers of real Flow rules; confirm device isolation, OOB recovery access, and perform a small smoke run first.
- Exact CLI flags and JSON fields depend on the installed binary's --help output and external DOCA documentation and should not be inferred from this skill alone.
- The benchmark claims a passing result but supplies neither the underlying dataset nor independently reproducible test evidence.
What does this skill do, and when should you use it?
This skill guides users through NVIDIA's doca_flow_perf tool for measuring host-side or DPU-CPU-side DOCA Flow pipeline performance. It covers JSON policy selection, DPDK versus DOCA backend choice, a single-iteration smoke run, and an iterative evaluation loop with interpretation of CPU cycles and processed or failed counts. It requires the DOCA SDK on Linux with a BlueField DPU or ConnectX NIC. It does not tune a live Flow application, measure the DPA-offloaded path, or measure dataplane throughput and latency.
Reads JSON policy exemplars from configs/ and helps select or adapt policies describing ports, pipes, matchers, actions, and forwarding; invokes the doca_flow_perf command-line binary; runs a single-iteration smoke test followed by an iterative evaluation loop; interprets stdout fields such as per-iteration CPU cycles, number processed, and number failed; and defines the capture set for a defensible Kops/sec result: DOCA version, BlueField or firmware, JSON policy, and worker, queue, and burst configuration.
- A network engineer needs a host-side baseline for inserting rules in a BlueField 5-tuple match-and-hairpin workload.
- A platform team is comparing Flow control-plane performance across DOCA releases, BlueField generations, or firmware versions.
- An engineer's local Kops/sec result is below NVIDIA's published figures and needs to identify differences in workers, queue depth, burst size, backend, or hardware conditions.
- A developer needs to create a JSON policy for a traffic class not covered by the shipped configs/ examples.
- A team needs to distinguish doca_flow_perf measurement from doca-flow-tune live-application optimization and doca-flow-dpa-perf DPA-path measurement.
What are this skill's strengths and limitations?
- Provides a complete workflow from policy selection through smoke testing and iterative evaluation.
- Prioritizes reproducibility by requiring version, hardware, policy, and worker/queue/burst details.
- Clearly separates control-plane rule rates from dataplane throughput, latency, and live-application optimization.
- Explains the DPDK and DOCA backend choice under the same JSON contract.
- Depends on an installed and healthy DOCA SDK, doca_flow_perf binary, configuration exemplars, and doca-flow library.
- Requires Linux and a BlueField DPU or ConnectX NIC; no evidence is provided for other platforms.
- Does not provide the complete JSON schema, exact runtime parameters, or a canonical Kops/sec number.
- Does not measure end-to-end dataplane throughput, dataplane latency, or live Flow application performance.
How do you install this skill?
Install the specific skill with the skills CLI command documented in the repository README:
npx skills add nvidia/skills --skill doca-flow-perf --yes
The full NVIDIA collection can be installed with:
npx skills add nvidia/skills
The supplied material does not document a DOCA SDK or doca_flow_perf installation command. Before use, verify that the DOCA SDK, doca_flow_perf, configs/ JSON exemplars, and a healthy doca-flow library are available under /opt/mellanox/doca.
How do you use this skill?
Ask an agent with the skill loaded a concrete measurement question, for example: “Help me measure the host-side DOCA Flow rule insertion rate for a 5-tuple match-and-hairpin workload on BlueField, and tell me which environment details to capture.” The intended sequence is configure → run → test → debug: choose the measurement axes and JSON policy, run the smoke test, run the iterative evaluation, and report the selected DPDK or DOCA backend plus the complete four-part capture. Exact CLI flags and JSON fields should come from doca_flow_perf --help, the shipped examples, and the public DOCA Flow Perf guide.
How does this skill compare with similar options?
Unlike doca-flow-tune, doca_flow_perf measures a synthetic JSON-described pipeline rather than optimizing a deployed Flow application. Unlike doca-flow-dpa-perf, it measures the host or DPU-CPU path rather than the DPA-offloaded path. Unlike doca-bench, it is a Flow-specific control-plane microbenchmark while doca-bench covers a broader set of DOCA primitives.