DOCA PCC Diagnostic Counters
Read fixed firmware and hardware PCC congestion diagnostics on ConnectX and BlueField devices.
The skill explicitly discloses sudo/debugfs writes, device scope, high-stakes tuning boundaries, lack of credential use, and the separation from custom algorithms, while requiring target-device checks. It lacks an explicit per-operation user-confirmation step and a concrete rollback or recovery procedure, so points are deducted.
The two-argument contract, set-then-query order, and layered error taxonomy are reasonably consistent and include diagnostic guidance. However, the supplied material does not include the actual script or a statically verifiable test suite covering key paths, so the static ceiling applies and points are deducted.
Audience, trigger phrases, use cases, and non-fit boundaries are clearly stated, with Linux/DOCA/MFT/hardware prerequisites. There is no evidence of Chinese-language support or mainland-China reachability, and the skill depends on specialized hardware, so points are deducted.
The documentation is layered across SKILL.md, CAPABILITIES.md, and TASKS.md and includes prerequisites, examples, limitations, routing, and error taxonomy. Versioning, authorship, changelog, and maintenance ownership are incomplete; the skill card still has TBD version and signature fields, and license metadata is inconsistent between Apache-2.0 and Apache-2.0 AND CC-BY-4.0.
The core task, command contract, fixed counters, and output format are described directly enough to guide set/query execution. However, the skill does not ship the script, actual results must come from the target device, and no representative output is verifiable in this review, so only a moderate static score is justified.
The material provides source references, concrete symbols, and a claimed passing benchmark report. The benchmark dataset is unavailable, the report is not independently reproducible evidence, and the actual script and coverage tests are absent, leaving a thin evidence chain.
- Before running set, reconfirm the exact MST device path and explicitly disclose that it is a root-level debugfs write.
- No rollback or recovery procedure is supplied for set; do not directly convert counter readings into congestion-control parameter changes.
- The benchmark report lacks its underlying dataset, and the actual script is absent from the supplied material; its pass rates are not independently reproducible here.
- Chinese-language support and mainland-China reachability are not demonstrated; linked public documentation may be inaccessible in restricted networks.
What does this skill do, and when should you use it?
This tool skill invokes NVIDIA's pcc_counters.sh script to arm and read fixed Programmable Congestion Control diagnostic counters. It uses mst and the mlx5 debugfs diag_cnt interface to report CNP, RTT, WRED-drop, and related event counters. It is intended for operators, developers, and AI agents collecting device-level evidence during congestion-control investigations. It does not write custom PCC algorithms, configure the factory firmware algorithm, or install DOCA/MFT.
The set operation writes counter IDs and parameters to arm the device diagnostics. The query operation reads /sys/kernel/debug/mlx5/<pci>/diag_cnt/dump and prints named counters. The script resolves an mst device path to a PCI address using mst status -v and lspci. Its fixed counter families include PCC_CNP_COUNT, MAD_RTT_PERF_CONT_*, *_EVENT_WRED_DROP, HANDLED_*_EVENTS, DROP_RTT_PORT*, and RTT_GEN_PORT*.
- A network operator checks whether a ConnectX or BlueField device is producing CNP, RTT, or WRED-drop events.
- A developer correlates a custom doca-pcc algorithm with device-level PCC diagnostic counters.
- An operator investigates whether a zero counter reflects idle traffic, missing arming, or genuinely absent events.
- An AI agent captures a PCC counter snapshot as evidence for a congestion-control investigation.
- An administrator diagnoses a Bad Device error, missing debugfs mount, or insufficient permissions.
What are this skill's strengths and limitations?
- Reads device firmware and hardware PCC diagnostics independently of a custom doca-pcc kernel.
- Covers CNP, RTT, WRED-drop, and handled-event diagnostics.
- Documents layered failure modes including bad-device resolution, missing arming, debugfs, permissions, and zero counters.
- Requires no compilation; the shipped surface is a Bash script.
- Requires Linux, DOCA/MFT, mst tools, mounted debugfs, and root access to mlx5 diag_cnt.
- Offers only set and query with two positional arguments; it has no watch, diff, or snapshot mode.
- Values are device-, firmware-, and traffic-state-specific, so example values are not portable evidence.
- Does not teach custom PCC algorithm development, firmware PCC configuration, or DOCA installation.
- The source does not provide a detailed platform test matrix or standalone test suite.
How do you install this skill?
Install the skill from the NVIDIA/skills collection with the documented skills CLI flow: npx skills add nvidia/skills --skill doca-pcc-counters --yes. The source does not document the script's exact absolute installation path. The target system must also have DOCA/MFT installed and meet the Linux, mst, debugfs, and root-access prerequisites.
How do you use this skill?
On a prepared host or BlueField Arm system, run set before query against the mst device: pcc_counters.sh set /dev/mst/mt41692_pciconf0; pcc_counters.sh query /dev/mst/mt41692_pciconf0. Use the set-then-query sequence when collecting counters. The script has exactly two operations and does not provide --help, --version, list, snapshot, watch, or diff.
How does this skill compare with similar options?
This is separate from doca-pcc: this skill reads device-level firmware and hardware counters, while doca-pcc writes and loads custom congestion-control kernels. It is also separate from firmware PCC algorithm configuration and the doca-setup installation workflow.