DOCA Flow DPA Performance
Guides defensible measurement of DOCA Flow rule-update and disable rates on DPA-capable NVIDIA hardware.
The files constrain supported hardware, VNF mode, PF/VF usage, require user confirmation, recommend a smoke run, warn against production use, and disclose device binding, traffic verification, and version dependencies. Rollback and recovery are not adequately specified, while the tool binds devices, allocates resources, and programs path-selector entries through an install-dependent environment, so points are deducted.
TASKS.md and CAPABILITIES.md provide configuration, smoke, measurement, debugging, and layered failure guidance. However, this review is static, key commands were not reproduced, exact parameters depend on the installed README and --help, and no executable skill-local test suite is supplied; the static ceiling therefore applies and points are deducted.
Audience, supported hardware, active/passive topology, workload axes, and out-of-scope boundaries are clearly stated, with both explicit and implicit trigger examples. Chinese-language support and mainland-China reachability are not demonstrated, and the core workflow depends on a local installation and external documentation, so points are deducted.
The SKILL.md, CAPABILITIES.md, and TASKS.md provide useful layering, examples, routing, limitations, dependency notes, and safety rules. Stable release versioning, changelog, and explicit maintenance ownership are incomplete; SKILLCARD version/signature fields remain TBD, the benchmark reports schema findings, and license metadata is inconsistent, so points are deducted.
The skill usefully structures DPA performance measurement around prerequisites, workload selection, smoke testing, statistics interpretation, and failure routing, which can reduce invalid measurements. It supplies no runnable scripts, fixed output, or reproducible fixture, and results remain dependent on hardware, firmware, DOCA version, and external README content; static evidence is insufficient for a higher score.
The material includes internal cross-references, a four-tuple capture rule, iteration statistics, optional tcpdump self-test, and an evaluation report. The evaluation dataset is unavailable, no real CI plus committed key-path test suite is shown, and this review did not execute anything, so only limited static auditability is supported.
- This is a low-confidence static review; doca_flow_dpa_perf and its self-test were not executed.
- Before running, verify the installed README and --help rather than relying only on workload defaults or parameter descriptions in the skill.
- The tool touches hardware and may affect traffic; confirm a non-production environment, bounded device targets, and a recoverable operating plan.
- Chinese-language usability and mainland-China network reachability are not demonstrated; external DOCA documentation may be an operational blocker.
- The SKILLCARD version, signature, and timestamp remain TBD, and license information is inconsistent; correct these before publication.
What does this skill do, and when should you use it?
This tool-oriented Agent Skill guides invocation of doca_flow_dpa_perf for the DPA-offloaded DOCA Flow update path. It helps operators verify hardware, VNF Flow mode, device roles, and workload-shape parameters before measuring. It also explains iteration statistics, self-test output, Kops/sec reporting, and layered failure diagnosis. It is not intended for the host or DPU-CPU Flow path, Flow tuning, DOCA installation, or DOCA API development.
Guides the agent to inspect the DOCA installation under /opt/mellanox/doca, pkg-config doca-flow, and the installed doca_flow_dpa_perf binary and README; verify ConnectX-7+ or BlueField-3 hardware, VNF Flow mode, and PF/VF constraints; select burst, queue, completion-threshold, worker, hash-pipe, work-policy, and PSL-table parameters; run a small smoke test followed by an evaluation loop; and interpret update or disable-enable iteration statistics, self-test results, and optional tcpdump verification.
- A platform operator deciding whether a path-selector workload should move to DPA or remain on the host or DPU-CPU Flow path.
- A performance engineer creating a device-, DOCA-version-, and workload-scoped DPA Kops/sec baseline.
- A DOCA Flow developer measuring the deployed performance of a DPA-offloaded rule-update path.
- An operator diagnosing zero-operation, hung, or failed-self-test runs through the documented error layers.
- An engineer reporting DPA and host Flow measurements together while keeping the producing tool and workload context explicit.
What are this skill's strengths and limitations?
- Draws a clear boundary between the DPA-offloaded path and the host or DPU-CPU Flow path.
- Covers hardware prerequisites, workload axes, iteration statistics, self-testing, and error classification.
- Promotes smoke-before-bulk testing, four-tuple capture, and explicit tool naming for safer comparisons.
- Explicitly identifies BlueField-2 and earlier ConnectX devices as unsupported.
- Requires an installed DOCA SDK, matching binary, DPA-capable hardware, and device permissions.
- Ships guidance rather than scripts, parsers, samples, or fixed throughput benchmarks.
- The supplied material omits the complete CLI flag inventory, defaults, and reproducible performance baselines.
- Does not cover DOCA installation, Flow API programming, or host-side Flow performance testing.
How do you install this skill?
Install the selected skill from NVIDIA/skills with: npx skills add nvidia/skills --skill doca-flow-dpa-perf --yes. The CLI prompts for an installation destination; cloning the repository or manually copying the folder is not required. DOCA, hardware, and device-configuration prerequisites are separate from skill installation.
How do you use this skill?
After installation, ask the agent a task-specific question such as “Measure the DPA-offloaded Flow rule-update rate on ConnectX-8.” The agent should verify DOCA, hardware, and VNF Flow prerequisites, then follow the configure, run, and test workflow in TASKS.md and use CAPABILITIES.md to interpret results. The supplied material does not provide a complete CLI invocation or default-value inventory; use the installed README or doca_flow_dpa_perf --help for those details.
How does this skill compare with similar options?
The skill names doca-flow-perf as the appropriate alternative for the host or DPU-CPU Flow performance path and positions doca-flow-tune as a subsequent Flow-tuning tool. Results from the two performance surfaces should not be treated as interchangeable without naming the producing tool.