What does this skill do, and when should you use it?
This tool skill belongs to NVIDIA’s Agent Skills catalog and targets developers and platform operators who already launch a DPA workload through doca-dpa. It guides doca_dpa_hl_tracer capture of higher-level DPA events, including kernel entry and exit, synchronization points, communication primitives, RDMA work-request submission, and completion draining. It covers TRACE versus CRIT, JSON configuration, file limits, thread priorities and CPU affinity, and decoding with the matching DPA-side ELF. It is not intended for writing DPA kernels, host-side debugging, raw cycle profiling, or production telemetry.
Checks prerequisites involving the DOCA installation, BlueField DPA visibility, DPACC-built ELF, and a running workload; selects TRACE or CRIT capture; tunes log and binary file limits, file-size-limit policy, thread priorities, and core affinity; runs doca_dpa_hl_tracer to produce a binary trace file and readable log; decodes the capture against the matching DPA-side ELF; and diagnoses empty, noisy, truncated, or undecodable captures across installation, device binding, image instrumentation, capture window, decoding, version, and overhead layers.
- A DPA developer whose kernel reports a wrong result while host-side completions look clean, using DPA-side events as runtime evidence.
- An engineer investigating excessive latency from kernel entry to the first communication call or from RDMA WR submission to completion drain.
- A platform operator whose TRACE capture materially increases latency and needs a lower-overhead capture strategy.
- An operator whose binary trace reaches a size limit or stops mid-run and needs to tune file limits and the limit policy.
- A user whose tracer writes a file but the parser reports zero events and needs to distinguish missing events, ELF mismatch, and installation problems.
How do you install this skill?
- Do not treat the benchmark PASS result or NVIDIA provenance as independent proof of reliability, effectiveness, or safety; the report states that the source evaluation dataset was unavailable.
- Before execution, verify the installed binary's --help, required privileges, mode tokens, JSON keys, and exact ELF match; trace files may contain sensitive workload information.
- Resolve the license metadata conflict between SKILL.md and SKILLCARD.yaml, and add versioning, changelog information, and reproducible key-path test coverage.
- Shell / CLI
- Local filesystem
DOCA SDKDOCA DPA ToolsBlueField device with an exposed DPA processorDPACC-built DPA-side ELFlive doca-dpa-launched workloadLinux: Ubuntu 22.04/24.04 or RHEL/SLES
Install the individual skill from the NVIDIA/skills catalog with the skills CLI:
npx skills add nvidia/skills --skill doca-dpa-hl-tracer --yes
This installs the skill directory; it does not install DOCA. Actual use also requires the DOCA SDK and optional DOCA DPA Tools on Linux, an accessible BlueField DPA processor, a matching DPACC-built ELF, and a live doca-dpa-launched workload.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/doca-dpa-hl-tracer" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Use doca_dpa_hl_tracer to investigate the latency between DPA kernel entry and the first RDMA WR submission, and check whether the trace was truncated.
Confirm that the question concerns DPA-side high-level tracing, then configure the capture window and JSON configuration. Start with CRIT to limit observation overhead and widen to TRACE when per-event detail is required. The documented invocation shape is:
doca_dpa_hl_tracer --mode <TRACE-or-CRIT> --config-file <config-file> --output-file <output-file>
Decode the resulting capture with the ELF that matches the loaded DPA image. The supplied source does not include a complete JSON example, decode command, or full parameter inventory, so use the installed binary’s --help and the applicable DOCA DPA Tools documentation. Example trigger: “Use doca_dpa_hl_tracer to investigate the latency between DPA kernel entry and the first RDMA WR submission, and check whether the trace was truncated.”
What are this skill's strengths and limitations?
- Provides a focused capture, decode, and troubleshooting workflow for DPA programming events.
- Explains the fidelity-versus-overhead tradeoff between TRACE and CRIT.
- Keeps DPA-side tracing separate from host-side doca-dpa control and raw cycle profiling.
- Addresses practical failure points including file limits, capture windows, ELF matching, and version alignment.
- Requires an existing DOCA installation, DOCA DPA Tools, BlueField DPA, matching ELF, and a running workload.
- Ships no sample traces, parsers, wrapper scripts, or samples/reference subtree.
- The supplied material omits a complete JSON example, decode command, and full event inventory; those must come from --help and the public documentation.
- Tracing is bounded debugging evidence, not a production observability solution.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| DOCA DPA High-Level Tracer this page ✓ NVIDIA · Official | 50 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA Environment Setup ✓ NVIDIA · Official | 57 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA Telemetry Exporter Development ✓ NVIDIA · Official | 55 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA Programming Guide ✓ NVIDIA · Official | 54 · Use with care | ★ 3.8k | today | Apache-2.0 |
| DOCA Version Consistency Assistant ✓ NVIDIA · Official | 53 · Use with care | ★ 3.8k | today | Apache-2.0 |
The source explicitly distinguishes this high-level tracer from raw cycle-level profiling, host-side doca-dpa debugging, and DPA kernel programming; those follow different skills or tool paths.
How did FollowSkills review this skill?
The documentation defines a bounded diagnostic tool, warns against continuous production use, requires retaining configuration and ELF identity, and discloses that traces may contain workload-sensitive RDMA data. It does not provide concrete least-privilege guidance, user-confirmation gates, rollback procedures, or redaction steps, so points are deducted.
The workflow, layered error taxonomy, and failure feedback are fairly detailed, and versioned parameters are explicitly delegated to --help and live documentation. However, static review cannot reproduce the key path, commands depend on external version-specific surfaces, and TASKS.md ends with an incomplete item, so the score remains conservative.
Trigger examples, intended scenarios, and out-of-scope boundaries are unusually clear for this tool. Chinese-language interaction and mainland-China reachability are not addressed, while the core workflow depends on external NVIDIA documentation, DOCA, BlueField hardware, and DPACC, so points are deducted.
The skill has strong information architecture, loading guidance, dependencies, limitations, related-skill routing, and issue contacts. However, version, signature, and scan fields remain TBD; the SKILL.md Apache-2.0 license conflicts with SKILLCARD.yaml's Apache-2.0 AND CC-BY-4.0 claim; no clear changelog is supplied; and the benchmark report records schema and recommended-section findings.
It offers a directly actionable configuration, capture, decode, diagnosis, and handoff workflow with clear marginal value for DPA tracing. It ships no scripts, sample traces, or directly verifiable outputs, and still relies on the installed binary's --help and external documentation, so static calibration limits the score to 6.
The files provide detailed taxonomy and parameter claims plus a benchmark report, but the benchmark dataset is unavailable and no independently reproducible test suite or real CI coverage of the skill's key paths is supplied. Public sources are pointers rather than attached corroborating evidence, so verifiability is limited.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →