DOCA Bench Benchmarking Skill
Measure DOCA library throughput, latency, and bandwidth reproducibly on real NVIDIA networking hardware.
The documentation defines a non-production warning, cautions about sensitive information on the companion-app channel, requires smoke-before-bulk execution, and forbids invented parameters. Deductions apply because least privilege, explicit user confirmation, rollback, and data-redaction procedures are not specified; requiring a full unredacted summary and CSV may expose sensitive environment details.
The configure, run, test, and layered-debug workflows are reasonably coherent and cover version, device, granular-build, warm-up, steady state, and failure routing. Deductions apply because no key path was executed in this static review, failure feedback depends on external documentation and installed --help, and TASKS.md inconsistently describes three modes while other files define four.
Trigger conditions, audience, use cases, and non-fit boundaries are clearly stated, with host and BlueField Arm coverage and install-side capability probing. Deductions apply because Chinese-language support and mainland-China reachability are not addressed, core references depend on docs.nvidia.com, and platform/library support boundaries require external matrix verification.
The skill uses layered SKILL.md, CAPABILITIES.md, and TASKS.md documentation with loading order, related skills, limitations, licensing, and routing guidance. Deductions apply because author, release version, signature, changelog, and maintenance responsibility remain TBD or indirect in SKILLCARD.yaml; the skill-card license conflicts with frontmatter, and static validation reported structural and recommended-section findings.
The skill clearly guides library, workload, and measurement-mode selection, baseline capture, smoke-before-bulk execution, and iterative diagnosis. Deductions apply because it provides no complete copy-ready commands or verified sample outputs, relies on the installed binary and external guide, and static evidence cannot establish usable real-world results.
The material contains extensive internal cross-references, source-code symbols, and an evaluation report, providing some auditability. Deductions apply because the evaluation dataset is unavailable, the report cannot be independently reproduced, and there is no committed CI or test-suite evidence covering the skill's key paths.
- Confirm the target device is not serving production traffic, and apply redaction and access controls to the companion-app channel, CSV files, logs, and environment data.
- Do not use assumed parameters directly; verify them with doca_bench --help, the installed query results, and documentation matching the installed version.
- Resolve the documentation inconsistency between four benchmark modes and the TASKS.md statement describing three modes; also reconcile SKILLCARD license, version, and signature metadata.
- Users in mainland China should verify reachability of docs.nvidia.com, NGC, and other external references before relying on the workflow.
What does this skill do, and when should you use it?
This skill guides AI agents in using doca_bench to microbenchmark RDMA, Compress, AES-GCM, SHA, DMA, EC, Ethernet, Comch, and GPUNetIO. It covers capability enumeration, three-axis configuration, smoke-before-bulk execution, stability checks, baseline capture, and layered failure diagnosis. It requires DOCA SDK 2.7.0 or newer on Linux with a BlueField DPU or ConnectX NIC. It does not cover application end-to-end timing, DOCA installation, custom benchmark code, or binary modification.
Inspects the installed DOCA environment and enumerable benchmark libraries; configures tests by target library, workload shape, and measurement axis; runs throughput, bulk-latency, precision-latency, or maximum-bandwidth measurements; performs a smoke-before-bulk workflow; captures commands, versions, devices, deployed environment, and screen/CSV results; and diagnoses zero, hanging, unstable, or unexpected results across configuration, device binding, library prerequisites, workload prerequisites, measurement soundness, version, and cross-cutting layers.
- A platform operator reruns a BlueField or ConnectX baseline before and after a firmware, driver, or NUMA change.
- A performance engineer compares actual DOCA Compress, RDMA, DMA, or SHA throughput or latency on a target device.
- A developer checks whether a candidate DOCA library is exposed and measurable on the current installation before designing an application.
- An SRE produces a citable performance artifact containing the command, version, device, environment, and measurements.
- An AI agent investigates zero output, startup hangs, or unstable doca_bench results.
What are this skill's strengths and limitations?
- Measures the user's actual installation and device instead of inferring performance from datasheet figures.
- Covers multiple DOCA libraries and four measurement axes.
- Requires recording command, version, device, and environment for regression analysis.
- Provides workflows for capability probing, measurement soundness, and layered diagnosis.
- Requires DOCA SDK >= 2.7.0, a suitable Linux environment, and NVIDIA hardware; it does not install DOCA.
- Ships no wrapper scripts, parsers, sample output, or expected throughput numbers.
- Flag and scenario details are installation-specific and should not be invented from the skill text.
- The source prompt provides no independent test-suite or platform-coverage evidence for this skill.
How do you install this skill?
Install the individual skill through the repository README's skills CLI flow:
npx skills add nvidia/skills --skill doca-bench --yes
Alternatively, run npx skills add nvidia/skills and select the skill and destination interactively. The repository is a catalog synchronized daily from product repositories; the skill becomes available the next time the agent loads relevant skills.
How do you use this skill?
On a Linux host meeting the prerequisites, or in an equivalent DOCA container, ask the agent for a doca_bench task such as: “Measure DOCA Compress throughput on my BlueField-3 and save a regression-testable baseline.” The workflow should verify the DOCA version and binary, enumerate enabled libraries, select throughput, bulk latency, precision latency, or maximum bandwidth, run a smoke test, and then assess stability. Remote-memory, RDMA, and Eth scenarios also require the companion app on the far side. Use the installed version's documentation and doca_bench --help for exact parameters.
How does this skill compare with similar options?
Unlike doca-caps, this skill performs fine-grained performance measurements while doca-caps provides a coarse per-device, per-library capability snapshot. Unlike application end-to-end benchmarking, it measures the DOCA library surface rather than the user's complete application. Unlike doca-setup, it assumes installation is already complete.