About doca-dpa-hl-tracer
doca-dpa-hl-tracer is an agent skill for invoking the DOCA DPA high-level tracing CLI. It guides developers and operators through capture, decoding, and rendering to investigate DPA behavior and performance while balancing event detail against tracing overhead.
Key features
- 01Capture DPA programming-event traces
- 02Choose between TRACE and CRIT modes
- 03Configure capture windows and file limits
- 04Decode binary traces into event streams
- 05Diagnose capture and decode issues by layer
Use cases
- 01Locate incorrect results or event ordering in a DPA kernel
- 02Analyze latency from kernel entry to communication calls
- 03Limit tracing overhead on a workload
- 04Investigate empty traces or decode failures
Install
Copy the command and run it in a terminal.
npx skills add https:/ / github.com/ nvidia/ skills --skill doca-dpa-hl-tracerAfter installing, type/doca-dpa-hl-tracerin Claude Code to confirm it loaded.
Bundled files
References and scripts installed along with SKILL.md.
BENCHMARK.mdCAPABILITIES.mdskill-card.mdskill.oms.sigSKILLCARD.yamlTASKS.mdevals/evals.jsonDetails
From the project READMEDOCA DPA High-Level Tracer
Examples
- "My DPA kernel is doing the wrong thing — where do I look?" —
this is in scope when there is no host-side
DOCA_ERROR_*. Capture then decode; do not start with cycle-only profiles. - "TRACE mode buried the DPA in observation overhead." — switch
capture window and
TRACEvsCRITbefore changing the kernel.
Instructions
Follow Loading order below. Open the tasks companion for configure / run / debug. Open the capabilities companion for event taxonomy. Do not open every companion file.
Where to start: This is a tool skill for invoking
doca_dpa_hl_tracer — the documented host-side CLI that
captures DPA-side execution traces in higher-level terms
(DPA programming events: kernel entry / exit, sync points,
comm primitive calls, RDMA WR submission, completions) rather
than raw cycle counts. Open the tasks companion and start at
## configure for the
mode-vs-overhead decision and the JSON config layout, then
## run for the
capture → decode → render pipeline. Open
the capabilities companion when the question is
what does this tool actually trace, which DPA programming
events does it expose, what is the trace-overhead vs
fidelity tradeoff, or how does it slot into a DPA debug
loop alongside doca-dpa
and doca-debug. If DPA is not
the right surface for the user's question (e.g. the bug is
host-side, the bug is in the DPACC-produced image, the user
wants raw cycle counts), the path-selection rule in
the capabilities companion
routes the agent before any capture is attempted.
Example questions this skill answers well
The CLASSES of doca_dpa_hl_tracer questions this skill is
built to answer, each with one worked example. The class is
the load-bearing piece; the worked example is one instance.
- "My DPA kernel is doing the wrong thing — where do I
look?" — worked example: "my host-side
doca_dpa_kernel_launch_update_*completes, but the kernel's reported result is wrong; no host-sideDOCA_ERROR_*". Answered by the when DPA-side high-level tracing is the right surface gate in the capabilities companion- the capture → decode → render flow in the tasks companion + the which DPA programming events to focus on rule in the tasks companion.
- "My DPA kernel is slow at a granularity that doesn't
show up in cycle profiles — how do I see kernel-entry to
first-comm-call latency?" — worked example: "my DPA
kernel runs but the time between launch and the first
RDMA WR submission is bigger than I expected". Answered
by the event-taxonomy table in
the capabilities companion
- the iterative loop in
the tasks companion which treats trace
overhead, mode (
TRACEvsCRIT), and capture window as axes to tune.
- the iterative loop in
the tasks companion which treats trace
overhead, mode (
- "How do I capture a trace without burying the DPA in
observation overhead?" — worked example: "
TRACEmode is producing too much data and my measured DPA latency went up by 2x compared to without the tracer". Answered by the mode-vs-overhead tradeoff in the capabilities companion- the
CRIT-first guidance in the tasks companion (start with critical-events-only; widen toTRACEonly when the bug demands per-event detail).
- the
- "My trace file got truncated mid-run — how should I
configure the file-size limits?" — worked example:
"binary trace file hit 5 GB and the capture stopped".
Answered by the
log_file_max_size_in_bytes/bin_file_max_size_in_bytes/file_size_limit_policytriple in the capabilities companion- the JSON config layout in the tasks companion.
- "Is the tracer on my install, and is it paired with the
matching
doca-dpalibrary and DPACC compiler version?" — worked example: "is the tracer ABI on my install compatible with the DPA image my DPACC just produced?". Answered by the version-overlay in the capabilities companion, which redirects to the canonicaldoca-versionchain and adds the tracer ↔doca-dpalibrary ↔ DPACC compiler match rule. - "The capture file looks empty / decode failed — is the
install broken, no events fired, or am I tracing the
wrong thing?" — worked example: "
doca_dpa_hl_tracerran, wrote a file, but the parser shows zero events". Answered by the layered error taxonomy in the capabilities companion (install / device-binding / DPA-image-instrumented / capture-window / decode-vs-elf / overhead-saturated / version / cross-cutting) + the layered walk in the tasks companion.
Audience
This skill serves external developers, platform operators,
and AI agents who have already brought up a DPA-side workload
through doca-dpa and now
need higher-level visibility into what the DPA kernel is
actually doing on the wire — DPA programming events
ordering, sync gaps, comm-call latencies, RDMA-WR / completion
timing — without dropping all the way down to raw cycle
counters. Concretely:
- A DPA developer who can launch their kernel cleanly from the host side but whose kernel's result is wrong or whose DPA-side performance is below expectation, and who needs a DPA-side ground truth before triaging.
- A platform operator running a DPA-using workload (RDMA
offload from accelerator, custom CC algorithm via
doca-pcc) and needs to localize a regression to the DPA side without instrumenting the application. - An AI agent producing a DPA-side trace report as
evidence for the host-side
doca-dpaladder when the host side reports clean completions but the DPA-side behaviour is wrong.
It is not for users debugging the tracer binary itself,
not a substitute for the live public DOCA DPA Tools
guide, not the right place for users learning how to
write a DPA kernel (that audience belongs in
doca-dpa plus the public
DOCA DPA / DPACC / DPA-Comms / DPA-Verbs guides), and not
the right place for raw per-instruction cycle profiling
(different surface, different tool — route via
doca-public-knowledge-map ## DOCA tools).
The tracer is shipped as a CLI binary under
/opt/mellanox/doca/tools/, not a library you link against.
The skill uses the same kind: tool three-file shape as
the rest of the bundle so the agent's task-verb contract is
uniform across libraries, services, and tools.
Language scope
doca_dpa_hl_tracer is a C++ host-side CLI. Its inputs are
its JSON config file, the DPA-side ELF (the
doca_dpa_app-class image produced by DPACC), and a running
DPA-side workload that the host-side doca-dpa lifecycle
already started. Its outputs are a binary trace file
(bin_file) and a human-readable log file (log_file).
The skill keeps the workflow guidance language-neutral —
the DPA-side workload it traces can be C compiled by DPACC
or any other DPA translation unit DPACC accepts — and
routes per-language questions to the public DPA / DPACC
guides via
doca-public-knowledge-map.
When to load this skill
Load this skill when the user is — or the agent needs to —
invoke doca_dpa_hl_tracer on a real host with DOCA
installed against a BlueField with a DPA processor visible to
the host, and the host-side
doca-dpa lifecycle has
already brought a DPA workload up at least once. Concretely:
- Capturing a DPA-side trace to localize a DPA kernel's
wrong-result or wrong-ordering behaviour when the
host-side
doca-dpalifecycle reports clean completions. - Capturing a DPA-side trace to localize a DPA-side performance gap (kernel-entry to first-comm latency, RDMA-WR-issue to completion gap, sync-point dwell time) at a granularity above raw cycle counts.
- Choosing between
TRACEandCRITcapture modes based on the bug-vs-overhead tradeoff and the available capture window. - Tuning the JSON config (thread priorities, core affinities, file size limits, file-size-limit policy) so the capture itself does not perturb the workload more than the bug it is investigating.
- Decoding a captured
bin_fileagainst the matching DPA-side ELF to render the human-readable event stream. - Capturing a side-effect-bounded trace as prerequisite
evidence for a host-side
doca-dpaladder step.
Do not load this skill for general DOCA orientation,
DPA-side programming model questions, raw cycle profiling,
or DOCA / DPACC install. For those, route to
doca-public-knowledge-map,
doca-dpa, or
doca-setup.
What this skill provides
This is a thin loader. Substantive material lives in two companion files:
- the capabilities companion — what
doca_dpa_hl_tracercaptures: the DPA programming event taxonomy (kernel entry / exit, sync points, comm primitive calls, RDMA WR submission and completion drain), the two documented capture modes (TRACEfor full per-event,CRITfor critical-events only), the trace-overhead-vs-fidelity tradeoff, the config-file shape (receiver / binary-writer / file-writer / printer threads with priority + core affinity, file size limits,file_size_limit_policy), the capture-window + workload-must-be-running invariant, the ELF-must-match-image rule for decode, the version-availability overlay (tracer ↔doca-dpalibrary ↔ DPACC compiler), the layered error taxonomy (install / device-binding / image-instrumented / capture-window / decode / overhead-saturated / version / cross-cutting), the observability surface (binary trace file + log file + tool's own stderr), and the safety policy (capture is bounded; tracing is not a production observability surface). - the tasks companion — step-by-step workflows for the in-scope
task verbs:
install(route to host-side DOCA install + DPA prerequisites),configure(mode + JSON config layout + capture window),build(route to install — the binary is shipped, the DPA-side application is user-built by DPACC),modify(refuse — do not patch the binary; modify the JSON config and the invocation instead),run(the capture flow with--mode,--config-file,--output-file),test(iterative loop tuning mode, window, and overhead),debug(walk the error taxonomy),use(consume the decoded trace in adoca-dpadebug session), plus aDeferred task verbsblock.
The skill assumes a host where DOCA is already installed at
the standard location, a BlueField with a DPA processor is
present and visible to the host, the DPACC compiler is
installed at a version matched to the host-side DOCA, the
DPA-side application image (the ELF the tracer decodes
against) is on disk and matches what the
doca-dpa lifecycle loaded,
and the operator has the privileges the public DOCA DPA
Tools guide requires.
What this skill deliberately does not ship
This skill is agent guidance, not a samples or scripts bundle. To keep the boundary clean, it deliberately does not contain — and pull requests should not add:
- Specific flag strings, event names, or mode tokens
beyond what the public DOCA DPA Tools page and
--helpdocument. The DPA programming events surface evolves release to release;--helpon the installed binary is the authoritative inventory. - Pre-baked example traces or expected event timings. Trace output is workload-, DPA-image-, BlueField-, and firmware-specific; a captured example pinned to one setup misleads operators elsewhere.
- Wrappers, parsers, or rendering scripts in any language that consume the binary trace format. The format is documented; users who want to script against it should read the live guide and write the parser against their installed version.
- A specific tuning recommendation derived from a single trace. A DPA-side perf decision (move a sync, batch a comm call, change a launch argument) is a workload question and the skill prescribes how to capture and read traces — it refuses to translate a captured gap into a kernel-rewrite recommendation without the user's own analysis.
- A
samples/orreference/subtree. This is a thin loader for a shipped CLI; substantive material lives on the public page, in--help, and indoca-dpa.
Loading order
- Read this skill first to confirm the user's question is in scope (DPA-side high-level tracing, not DPA-side programming and not raw cycle profiling).
- For the event taxonomy, capture modes, overhead tradeoff, JSON config layout, version overlay, error taxonomy, observability, and safety policy, see the capabilities companion.
- For the documented invocations and the
capture → decode → render workflow —
install,configure,build,modify,run,test,debug,use— see the tasks companion.
Related skills
doca-dpa— the host-side DPA control library whose loaded application image the tracer captures. Pair them in every DPA debug session:doca-dpabrings the workload up; the tracer captures what the workload does at the DPA programming event layer. Conflating the library with the tracer is the most common DPA-debug first-touch error.doca-debug— the cross-cutting debug ladder. The tracer slots in at the runtime layer as the DPA-side ground truth before any DPA-side perf or correctness conclusion is made.doca-public-knowledge-map— routing to the public DOCA DPA Tools page ondocs.nvidia.comand the rest of the public DOCA documentation set.doca-version— canonical DOCA version-handling rules. The## Version compatibilitysection in the capabilities companion is a concise overlay that redirects here for the body and adds the tracer ↔doca-dpalibrary ↔ DPACC compiler matching rule.doca-setup— env preparation, install verification, DPACC compiler install / verification, BlueField mode (the DPA processor must be exposed before any tracing is meaningful), and the I have no install yet path with the public NGC DOCA container.doca-structured-tools-contract— the bundle's detect → prefer → fall back → report contract for structured helper tools. The command appendix in the tasks companion honors this contract.doca-programming-guide— general DOCA programming patterns shared by every library / tool surface, including the cross-libraryDOCA_ERROR_*taxonomy this tool's host-side error layer overlays on top of when host-sidedoca-dpacalls fail in tandem.
The DPA-side companion libraries doca-dpa-comms (comm
primitives the DPA kernel itself calls) and
doca-dpa-verbs (RDMA verbs the DPA kernel itself calls)
are different artifacts that the tracer's DPA
programming events surface visibly names; for the
DPA-side programming model itself, route through
doca-public-knowledge-map
to the public DOCA DPA-Comms and DPA-Verbs guides and to
the shipped /opt/mellanox/doca/samples/doca_dpa/ samples.
This tool traces their use; it does not redefine them.