DOCA SPCX Programmable Congestion Control
Operate and evaluate SPCX congestion-control algorithms on BlueField DPA before production rollout.
The documents explicitly require least-exposure production rollout, replica-first validation, out-of-band management, bounded deployment, rollback to factory PCC, and disclosure of fabric, firmware, algorithm, and parameter context. However, user confirmation, privilege boundaries, concrete rollback commands, and sensitive-data handling remain incomplete for a live congestion-control tool, so 8 points are deducted.
SKILL.md, CAPABILITIES.md, and TASKS.md are broadly consistent on scope, prerequisite checks, error taxonomy, and failure routing. However, many CLI details are deferred to the installed --help and public guides, with no reproducible command set, complete failure-output examples, or key-path test coverage. Under static calibration the score cannot exceed 10, so 12 points are deducted.
The audience, example scenarios, exclusions, semantic trigger phrases, SPCX/PCC/factory-PCC routing, and contention prerequisite are clearly described. However, the skill requires specific BlueField, DOCA, DPACC, and controllable RDMA-fabric conditions and provides no evidence of Chinese-language support or reachability from mainland-China networks, so 3 points are deducted.
The entry, capability, task, and benchmark files provide progressive disclosure, related-skill routing, limitations, and safety policy. However, versioning, changelog, maintenance ownership, and signing remain TBD or unclear; SKILLCARD repository and license metadata conflict with the supplied object, and the benchmark report says its source dataset was unavailable. Eight points are deducted.
For operators who already have compatible hardware, installation, and an algorithm image, the decision tree, prepare-smoke-contention workflow, troubleshooting ladder, and rollout gate address the core task. However, there are no bundled scripts, fixed invocation recipes, representative outputs, or executable tests; results depend on external documentation and live systems. Under static calibration the score cannot exceed 7, so 9 points are deducted.
The source provides substantial internal cross-references, an error taxonomy, evidence-field requirements, and a benchmark report. However, the benchmark dataset is unavailable, its execution claims cannot be independently reproduced from the supplied material, and there is no skill-specific executable test or CI evidence. Under static calibration the score cannot exceed 5, so 6 points are deducted.
- This is guidance for a high-risk real-time congestion-control tool; do not load it directly on production fabric until replica validation, observability, OOB access, bounded rollout, and rollback rehearsal are complete.
- CLI flags, roles, probe formats, and metrics may vary by DOCA release; verify them against the installed doca_spcx_cc --help and official guide rather than treating the prose as a complete interface.
- The supplied material does not establish mainland-China reachability or provide Chinese operating material; dependent public documentation or external services require separate validation.
- The SKILLCARD contains inconsistent license and repository metadata, and its version and signature fields are TBD; maintainers should correct these before publication.
What does this skill do, and when should you use it?
This skill helps external developers, platform operators, and AI agents use the host-side doca_spcx_cc CLI to manage an SPCX programmable congestion-control algorithm on a BlueField DPA. It covers loading, parameterization, starting, observing, stopping, and choosing SPCX versus doca-pcc or factory firmware PCC. Evaluation requires a live RDMA/RoCE link with controllable contention, and the workflow emphasizes replica-first testing, evidence collection, and rehearsed rollback. It does not author DPA-side algorithms, configure factory PCC firmware, provide general DOCA installation guidance, or replace low-level binary debugging.
Guides checks for the DOCA installation, BlueField DPA exposure, firmware custom-PCC slot, and DPACC version alignment; uses doca_spcx_cc under /opt/mellanox/doca/tools/ to load, parameterize, start, observe, and stop an SPCX algorithm; applies the SPCX-versus-doca-pcc decision tree; runs prepare, smoke, and contention-positive evaluation on a non-production RDMA/RoCE fabric; classifies failures across installation, device binding, firmware slot, DPA image, algorithm prerequisites, live-link prerequisites, runtime, and version layers; and assesses production-forward decisions using observability evidence, bounded blast radius, and rollback planning.
- A congestion-control researcher has a DPACC-compiled SPCX algorithm and needs host-side loading and runtime control on BlueField.
- A platform operator wants to create controlled contention on a non-production RoCE fabric and compare a custom SPCX algorithm with factory PCC behavior.
- A developer sees a clean load but unchanged throughput or latency and needs to determine whether the link lacks contention or factory control is still active.
- An infrastructure team needs an evidence-based decision on whether a tested SPCX algorithm can move toward production.
- An operator receives DOCA_PCC_PS_ERROR and needs to separate installation, device, firmware, DPA-image, version, and runtime causes.
What are this skill's strengths and limitations?
- Provides routing between SPCX, established doca-pcc, and factory firmware PCC surfaces.
- Builds in replica-first evaluation, contention requirements, runtime observability, and rollback gates.
- Offers a layered failure taxonomy and version-alignment checks useful for operations.
- Clearly separates host-side tool operation from DPA-side algorithm authoring.
- Requires a specific DOCA SDK, BlueField DPA, firmware custom-PCC slot, and compatible DPACC compiler.
- Requires a live RDMA/RoCE fabric with actual contention; without contention the algorithm may have no visible signal.
- Ships no congestion-control algorithm, example output, wrappers, parsers, scripts, or tuning recommendation.
- The supplied SKILL.md omits concrete flags and subcommands; installed --help is authoritative.
- It is out of scope for factory PCC configuration, general DOCA setup, DPA algorithm development, and raw DPA cycle profiling.
How do you install this skill?
Install the selected skill with the README-supported skills CLI command: npx skills add nvidia/skills --skill doca-spcx-cc --yes. It becomes available the next time the agent loads skills for a relevant task. The runtime environment must separately provide the DOCA SDK under /opt/mellanox/doca, the SPCX optional component, a BlueField exposing its DPA processor, firmware with the custom-PCC slot enabled, a version-matched DPACC compiler, and a non-production RDMA/RoCE fabric for evaluation.
How do you use this skill?
On a host meeting the hardware, firmware, software, and link prerequisites, ask for a task such as “I have a DPACC-compiled SPCX algorithm and a non-prod BlueField pair; how should I run a contention-positive evaluation?” or “doca_spcx_cc returns DOCA_PCC_PS_ERROR on start; how do I debug it by layer?”. The skill organizes guidance around tasks such as configure, run, test, debug, and use. Exact flags, subcommands, probe-packet tokens, and metric names must come from the installed doca_spcx_cc --help and public DOCA guides; the supplied source does not document a complete copy-paste invocation.
How does this skill compare with similar options?
The source defines three alternatives: SPCX is the newer programmable congestion-control surface; doca-pcc is the established host-side PCC path; and factory PCC is configured through firmware without doca_spcx_cc or a host-side library. The choice depends on installation support, algorithm availability, the firmware custom-PCC slot, DOCA/DPACC alignment, and whether the user is authoring or consuming an algorithm.