doca-telemetry — Claude Code skill: read DOCA hardware counters
Skill for reading DOCA hardware counters from BlueField DPU or ConnectX NIC via doca_telemetry_pcc, doca_telemetry_dpa, doca_telemetry_diag and other per-domain libraries.
Apache-2.0 ★ 4,106 updated 2026-10-23
npx skills add NVIDIA/skills --skill doca-telemetry -a claude-codeWhat doca-telemetry does
This skill covers the reading side of DOCA Telemetry — opening a per-domain context on doca_dev and reading current counter values across domains (PCC, DPA, DIAG, ADP_RETX, PHY, PCI). It walks you through capability checks, context creation, startup, and reading snapshots. This is not NetFlow, IPFIX, or a local-socket collector.
Publishing and exporting counters happens separately via doca-telemetry-exporter. For the deployed DOCA Telemetry Service (DTS), refer to doca-dts. The skill assumes DOCA SDK is already installed at /opt/mellanox/doca on Linux (Ubuntu 22.04/24.04 or RHEL/SLES) with a BlueField DPU or ConnectX attached. The library is read via pkg-config doca-telemetry.
Examples cover C and C++; other languages (Rust, Go, Python) call the same *.so through FFI or language bindings. The skill helps you choose the right per-domain header (doca_telemetry_pcc.h for PCC, doca_telemetry_dpa.h for DPA, and so on), check device support, and follow the per-domain lifecycle correctly.
What it helps with
- read PCC counters from a BlueField application
- sample DPA counter exports from a device
- expose PHY, PCI, or DIAG counters from a docadev
- check which counter domains your device supports
- open a per-domain context and manage its lifecycle
Example prompts for the agent
- Show me how to build code that reads PCC counters from my BlueField app using
doca_telemetry_pcc - How do I open a
doca_telemetry_dpacontext ondoca_devand read a counter snapshot? - Write an example that checks device support for DPA via
doca_telemetry_dpa_cap_is_supported, then reads the counters
How to install doca-telemetry
1Skills CLI (Vercel)
npx skills add NVIDIA/skills --skill doca-telemetry -a claude-code2Manually: put the files in place
~/.claude/skills/doca-telemetry/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/doca-telemetry/
1Skills CLI (Vercel)
npx skills add NVIDIA/skills --skill doca-telemetry -a codex2Manually: put the files in place
.agents/skills/doca-telemetry/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add NVIDIA/skills --skill doca-telemetry -a cursor2Manually: put the files in place
.cursor/skills/doca-telemetry/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install NVIDIA/skills skills/doca-telemetry2Manually: put the files in place
.github/skills/doca-telemetry/SKILL.md1Skills CLI (Vercel)
npx skills add NVIDIA/skills --skill doca-telemetry -a gemini-cli2Platform CLI
gemini skills install https://github.com/NVIDIA/skills --path skills/doca-telemetryManually: put the files in place
.devin/skills/doca-telemetry/SKILL.mdThe legacy .windsurf/skills/ path still works
Claude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about doca-telemetry
How do I install doca-telemetry in Claude Code?
Run npx skills add NVIDIA/skills --skill doca-telemetry -a claude-code, then open ~/.claude/skills/doca-telemetry/SKILL.md.
What does doca-telemetry do?
It is a library for reading DOCA hardware counters from a BlueField DPU or ConnectX NIC. The skill helps you open a per-domain context (doca_telemetry_pcc, doca_telemetry_dpa, and others) and read a counter snapshot.
What is the difference between doca-telemetry and doca-telemetry-exporter?
doca-telemetry is the reading side that fetches counters from the device. doca-telemetry-exporter publishes and exports them. Use the first when your code reads counters directly; use the second when you need to export them elsewhere.
Source text
SKILL.md · Apache-2.0
--- license: Apache-2.0 name: doca-telemetry description: > Use this skill to read DOCA hardware-counter events from a `doca_dev` through the per-domain Telemetry reader libraries: `doca_telemetry_pcc`, `_dpa`, `_diag`, `_adp_retx`, `_phy`, and `_pci`. It covers capability checks, context creation, startup, and per-domain reads or samples. Trigger for implicit requests such as "read PCC counters from my BlueField app", "sample DPA counter exports", or "expose PHY, PCI, or DIAG counters from this doca_dev". This is the counter-reader surface, not a NetFlow, IPFIX, or local-socket collector. Route publishing and export to `doca-telemetry-exporter`; route the deployed DOCA Telemetry Service to `doca-dts`, and collectors and plain stdout logging elsewhere. metadata: author: NVIDIA <[email protected]> kind: library doca_version: "3.5" tags: - doca - telemetry - counters compatibility: > Requires DOCA SDK installed at /opt/mellanox/doca on Linux (Ubuntu 22.04/24.04 or RHEL/SLES) with a BlueField DPU or ConnectX NIC attached. Reads the user's local install via `pkg-config doca-telemetry` and inspects /opt/mellanox/doca/{lib,include,samples,applications}. The validation steps require a live DOCA install. --- # DOCA Telemetry ## Examples - **"Build a second program that RECEIVES the events my exporter app publishes."** No DOCA library consumes exporter events, and this skill is not that library: `doca-telemetry` reads a device's hardware counters. The app links `doca-telemetry-exporter` to publish; the receiver is the DOCA Telemetry Service over IPC (`doca-dts`), or the exporter's file output for a local smoke. Do not invent a consume or collector API. Then: the publishing app owns the schema; check the IPC path with `doca_telemetry_exporter_check_ipc_status`; smoke one event with field values intact (file output) before bulk; when reporting cannot keep up, drop or aggregate on the hot path, never block. - **"Read PCC counters from my BlueField app."** — this is in scope. Use the per-domain Telemetry reader; do not stand up DTS or an exporter here. Sequence: `doca_telemetry_pcc_cap_is_supported` on the device, create the PCC context on the open `doca_dev` (`doca_telemetry_pcc_rep_create` for a representor), start it, read once with `doca_telemetry_pcc_get_counters`, then stop and destroy. Reads before start return `DOCA_ERROR_BAD_STATE`. - **"Sample DPA counter exports."** — stay on `doca_telemetry_dpa` reads. Publishing those samples belongs to the exporter skill. ## Instructions Follow Loading order below. Open the tasks companion for configure / read / debug. Open the capabilities companion for the per-domain readers. Do not open every companion file. **Where to start:** This skill assumes DOCA is already installed and the user is doing **hands-on hardware-counter-reader work** — opening a per-domain `doca_telemetry_<domain>` context against a `doca_dev` and reading the latest hardware-counter snapshot for that domain. The library is the **counter-READER** half of DOCA telemetry; it is NOT a NetFlow / IPFIX collector and NOT a generic schema-event consumer (there is no NetFlow / IPFIX / local-socket transport surface in the public header). Open the tasks companion if the user wants to *do* something (configure / build / modify / run / test / debug); open the capabilities companion when the question is *which hardware-counter domains can this device read* on this install (PCC, DPA, DIAG, ADP_RETX, PHY, PCI). If the user has not installed DOCA yet, route to `doca-setup/SKILL.md` first. If the user is confused about whether they want this library (HW-counter reader on a `doca_dev`) or `doca-telemetry-exporter/SKILL.md` (the publisher / export side, which is a separate library and publishes structured telemetry / labeled metrics / OTLP logs), read the reader-vs-exporter role split in the capabilities companion BEFORE configuring anything; mixing the two is the load-bearing first-app failure for this skill. If the user is asking about **DOCA Telemetry Service (DTS) as deployed**, route to `doca-dts`. ## Audience This skill serves **external developers building applications that READ DOCA hardware counters from a `doca_dev` through one or more of the six per-domain DOCA Telemetry reader libraries** (`doca_telemetry_pcc` / `_dpa` / `_diag` / `_adp_retx` / `_phy` / `_pci`) — i.e., users whose application code calls `doca_telemetry_<domain>_*` (directly in C/C++, or through FFI / bindings from another language) to open a per-domain context on a `doca_dev`, configure the per-domain sample window, and read the hardware-counter snapshot for that domain. It is *not* for NVIDIA developers contributing to DOCA Telemetry itself, and it is *not* for users writing the **publishing / export** side — that is `doca-telemetry-exporter/SKILL.md`, a separate library and a separate skill. **Language scope.** DOCA Telemetry's per-domain reader libraries ship as a C surface with `pkg-config` module name `doca-telemetry`. The shipped samples are written in C. C and C++ readers are the canonical case; the worked examples in the tasks companion assume that path. Other-language readers (Rust, Go, Python, …) consume the same `*.so` through FFI or language-specific bindings; the skill's contribution in that case is to keep the reader-vs-exporter distinction, the per- domain cap-query-first discipline, the per-domain DOCA Core lifecycle on the `doca_dev`, the sample-window discipline, and the error-taxonomy guidance language-neutral, and to route the agent to the public per-domain C ABI as the authoritative surface that any wrapper will eventually call. ## When to load this skill Load this skill when the user is doing hands-on DOCA Telemetry **hardware-counter-reader** work, in any language. Concretely: - Picking the right per-domain header for the counters the user wants (`doca_telemetry_pcc.h` for Programmable Congestion Control counters, `_dpa.h` for DPA counters, `_diag.h` for generic device diagnostic counters, `_adp_retx.h` for ADP retransmit counters, `_phy.h` for physical-layer counters, `_pci.h` for PCI / PCIe counters) and confirming the device supports it via the per-domain `_cap_is_supported(devinfo)` query — **except `_pci`, which has no single `_cap_is_supported` and instead exposes per-feature caps like `doca_telemetry_pci_cap_management_info_is_supported` / `_cap_perf_counters_1_is_supported`.** - Opening a per-domain `doca_telemetry_<domain>` context on a `doca_dev`, walking the per-domain lifecycle (`doca_telemetry_<domain>_create(dev)` → per-domain setters → `doca_telemetry_<domain>_start`), configuring the per-domain sample window, and reading the hardware-counter snapshot for that domain. Note this is a per-domain `_create`/`_start` surface, not the generic `doca_ctx_*` progress-engine lifecycle. - Reading the device + library capability surface before assuming a counter family is available: use `doca_telemetry_<domain>_cap_is_supported` only for `pcc`, `dpa`, `diag`, `adp_retx`, and `phy`; use the matching per-feature `doca_telemetry_pci_cap_*_is_supported` query for PCI. - Handling per-domain `DOCA_ERROR_*` returns from a counter read (lifecycle vs. device-doesn't-support-this-domain vs. per-domain `AGAIN`-means-snapshot-not-ready vs. permission / driver) and the per-read status reported back to the application. - Choosing between DOCA Telemetry (hardware-counter READER) and an adjacent option: `doca-telemetry-exporter/SKILL.md` when the user actually wants to PUBLISH / EXPORT the counter values (OTLP / Prometheus / labeled metrics); `doca-log` when plain structured stdout logging is enough; a generic Prometheus / OpenTelemetry client library when the counter source is a non-DOCA program; the DOCA Telemetry Service (`doca-dts`) when the user wants a turnkey aggregator. - Designing or extending non-C bindings (Rust, Go, Python, …) that wrap the per-domain reader C ABI — for the reader-vs- exporter distinction, the per-domain cap-query-first rule, the per-domain `doca_dev` lifecycle, the sample-window discipline, and the error rules the wrapper must honor. Do **not** load this skill for general DOCA orientation, install of DOCA itself, the **publishing / export** side (`doca-telemetry-exporter/SKILL.md` has its own skill), the DOCA Telemetry Service (`doca-dts`), or non-reader library questions. For those, use `doca-public-knowledge-map/SKILL.md`. ## What this skill provides This is a **thin loader**. The body keeps only the orientation needed to pick the right next file. The substantive hardware-counter-reader material lives in two companion files: - the capabilities companion — what the per-domain readers can express on this install: the reader-vs-exporter role-split rule, the six shipped sub-libraries (`doca_telemetry_pcc` / `_dpa` / `_diag` / `_adp_retx` / `_phy` / `_pci`) and which counter family each one exposes, the per-domain DOCA Core lifecycle on a `doca_dev`, the domain-level capability query for PCC / DPA / DIAG / ADP_RETX / PHY and the per-feature capability queries for PCI, the reader error taxonomy (mapped onto the cross-library `DOCA_ERROR_*` set, with the `NOT_SUPPORTED`-means-domain- not-exposed-on-this-device rule and the `AGAIN`-means- snapshot-not-ready rule called out explicitly), the observability surface (per-read status + per-domain cap- query snapshot at configure time), the safety policy that gates per-domain reads behind the cap-query result, and the path-selection rule against `doca-telemetry-exporter/SKILL.md`, `doca-log`, and standalone Prometheus / OpenTelemetry / DTS. - the tasks companion — step-by-step workflows for the six in-scope reader verbs: `configure`, `build`, `modify`, `run`, `test`, `debug`. Plus a `Deferred task verbs` block that points out-of-scope questions at the right next skill. The skill assumes a host where DOCA is already installed at the standard location and a target BlueField DPU or ConnectX NIC is available. The ## run workflow opens the corresponding `doca_dev` and requires the per-domain cap-query to return `DOCA_SUCCESS`. It does not cover installing DOCA — that path goes through `doca-setup/SKILL.md` — and it does not cover writing the publishing / export side, which is `doca-telemetry-exporter/SKILL.md`. ## Loading order 1. Read this skill first to confirm the user's question is in scope (specifically, that the user wants to READ a per-domain hardware counter via the per-domain reader API on a `doca_dev` — not PUBLISH counters, which is `doca-telemetry-exporter/SKILL.md`; not deploy DTS, which is `doca-dts`; and not stand up a NetFlow / IPFIX collector, which this library does not expose a surface for). 2. **For the reader-vs-exporter rule, the six per-domain sub-libraries, the per-domain DOCA Core lifecycle on a `doca_dev`, the per-domain capability query, the error taxonomy (including the `NOT_SUPPORTED`-means-domain- not-exposed-on-this-device rule and the `AGAIN`-means- snapshot-not-ready rule), observability, the safety policy, and the path-selection rule, see the capabilities companion.** 3. **For step-by-step workflows — configure, build, modify, run, test, debug — see the tasks companion.** Both companion files cross-link to each other, `doca-version/SKILL.md` for the canonical version-handling rules, and `doca-public-knowledge-map/SKILL.md` whenever the right answer is "look it up in the public docs or the installed package layout" rather than "reader-specific guidance". ## Related skills - `doca-telemetry-exporter/SKILL.md` — the **publishing / export** side. This skill (the per-domain hardware-counter reader) and `doca-telemetry-exporter` (the application-side labeled-metrics / OTLP-logs publisher) are two separate libraries that solve two different problems — they do NOT form a publisher/collector pair (there is no collector transport surface here). An application that wants to EMIT counters to its monitoring pipeline links the exporter; an application that wants to READ hardware counters off a `doca_dev` links this skill. Wiring the wrong library into the user's app is the #1 first-app failure for both skills, and the load-bearing rule the agent must surface before any code-level guidance. - `doca-public-knowledge-map/SKILL.md` — the routing table for every public DOCA documentation source and the on-disk layout of an installed DOCA package. The per-domain reader's public guide URL is `https://docs.nvidia.com/doca/sdk/DOCA-Telemetry/index.html` (distinct from the publisher's guide at `https://docs.nvidia.com/doca/sdk/DOCA-Telemetry-Exporter/index.html`); the on-disk samples live under `/opt/mellanox/doca/samples/doca_telemetry/`. The DOCA Telemetry Service (DTS) has its own skill, `doca-dts`. - `doca-setup/SKILL.md` — env preparation, install verification, and the *I have no install yet* path with the public NGC DOCA container. This skill assumes its preconditions are satisfied (in particular, a `doca_dev` is open and `doca_telemetry_<domain>_cap_is_supported(devinfo)` returns `DOCA_SUCCESS` for the per-domain library the user wants to read). - `doca-version/SKILL.md` — canonical DOCA version-handling rules. This skill's Version compatibility overlay cross-links the four-way match rule plus detection chain and adds the per-domain reader overlay (each per-domain library's counter set can grow across DOCA releases; cap-query first, then read). - `doca-structured-tools-contract/SKILL.md` — the bundle's structured-tools precedence rule (detect / prefer / fall back / report). The Command appendix in the tasks companion honors this contract. - `doca-programming-guide/SKILL.md` — general DOCA programming patterns shared by every library: the canonical `pkg-config` + meson build pattern, the universal modify-a-shipped-sample first-app workflow, the universal lifecycle, the cross-library `DOCA_ERROR_*` taxonomy, and the program-side debug order. This skill layers per-domain reader specifics on top. - `doca-log` — the right primitive when the user actually wanted plain structured stdout logging from their own program rather than a structured hardware-counter snapshot. The reader's per-domain cap- query discipline is overhead the user does not need for plain logs; this skill's path-selection rule routes there when log shipping is the actual requirement. - `doca-debug/SKILL.md` — the cross- cutting debug ladder (install / version / build / link / runtime / program / driver). Per-domain reader-specific debug (domain not supported on this device, snapshot not yet ready, sample window too short, cap-query forgotten before read) overlays on top of that ladder.
© NVIDIA · Apache-2.0 · NVIDIA/skills