What does this skill do, and when should you use it?
This skill guides operators in using DOCA Socket Relay to connect a host-side socket application with a peer on a BlueField DPU. It covers in-process, sidecar, and BlueField service-container deployment shapes, along with host socket and DPU forwarding-endpoint configuration. It provides a bind-to-connect, round-trip, and smoke-before-fleet workflow, plus layered diagnosis for connection failures and silent data paths. It assumes DOCA is already installed and routes comch programming, raw packet I/O, and DOCA installation to other skills.
Reads TASKS.md and CAPABILITIES.md to select deployment shape, socket type, and forwarding endpoint; guides configuration of the host-side listening socket and DPU-side endpoint; walks through binding, connecting, an end-to-end round trip, and admitting a client fleet only after a single-client smoke test; organizes diagnosis around missing installation, relay binding, host connection, DPU endpoint reachability, permissions, versions, and cross-cutting failures; and uses the installed version's --help, the public DOCA guide, and pkg-config version information for concrete command and compatibility decisions.
- A platform owner wants an existing host socket service to communicate with a peer on BlueField without rewriting the host application.
- A migration engineer wants to validate Socket Relay as an initial migration phase before moving to comch, RDMA, or Ethernet APIs.
- An SRE needs to choose between a host process, sidecar, and BlueField service-container deployment.
- An operator is diagnosing ECONNREFUSED, connection timeouts, accepted connections with no DPU-side data, or hangs after the first round trip.
- A platform team wants to validate one host client end to end before admitting the full client fleet.
How do you install this skill?
- Do not turn the conflicting connection-state inspection descriptions into commands; treat the installed version’s --help and actual logs as authoritative.
- Avoid retaining or forwarding unredacted host, DPU, or application logs; define sensitive-field filtering and least-privilege access first.
- State changes can cause disconnects, silent data-path failures, or partial data corruption; do not admit the full workload fleet before single-client, end-to-end, and rollback conditions are satisfied.
- Core operation depends on DOCA and BlueField hardware and may depend on external NVIDIA documentation; constrained or mainland-China environments should prepare local documentation and offline validation material.
- Shell / CLI
- Network access
- Local filesystem
DOCA SDK installed at /opt/mellanox/docaLinux with Ubuntu 22.04/24.04 or RHEL/SLESAttached BlueField DPUSocket-oriented host applicationBlueField kubelet-standalone runtime for container deployment
Use the skills CLI command documented in the repository README: npx skills add nvidia/skills --skill doca-socket-relay --yes. For a Codex-specific installation, use: npx skills add nvidia/skills --skill doca-socket-relay --agent codex --yes. The README says cloning the repository or copying skill folders manually is unnecessary; the skill becomes available when a relevant task is loaded.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/skills/doca-socket-relay" ~/.claude/skills/
rm -rf "$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:
- Forward my existing socket application to BlueField with a sidecar and verify one end-to-end round trip.
Ensure DOCA is installed and a BlueField DPU is attached, then describe the operational task in natural language, for example: “Forward my existing socket application to BlueField with a sidecar and verify one end-to-end round trip.” The agent should read SKILL.md first, then consult CAPABILITIES.md and TASKS.md as needed, beginning with configure and continuing with run, test, or debug. Binary names, flags, default paths, and ports are intentionally undocumented here; use the public DOCA Socket Relay guide and the installed binary's --help.
What are this skill's strengths and limitations?
- Targets BlueField migration for socket applications that should not be rewritten.
- Covers deployment choices, endpoint configuration, smoke validation, and layered troubleshooting.
- Clearly distinguishes Socket Relay's data-plane role from comch's control-plane role.
- The catalog is maintained through daily synchronization, signing, evaluation data, and benchmark reports.
- Requires an installed DOCA SDK and an attached, usable BlueField DPU.
- Does not include binary names, flag inventories, default ports, scripts, samples, or output parsers.
- It is agent guidance rather than a replacement for the public DOCA Socket Relay guide or installed help output.
- Container deployment additionally requires the BlueField kubelet-standalone runtime.
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 Socket Relay this page ✓ NVIDIA · Official | 49 · Use with care | ★ 3.6k | 13d ago | Apache-2.0 |
| DOCA Container Deployment ✓ NVIDIA · Official | 52 · Use with care | ★ 3.6k | 13d ago | Apache-2.0 |
| DOCA BlueField-3 Bring-Up ✓ NVIDIA · Official | 56 · Use with care | ★ 3.6k | 13d ago | Apache-2.0 |
| DOCA BlueField Bare-Metal Deployment ✓ NVIDIA · Official | 53 · Use with care | ★ 3.6k | 13d ago | Apache-2.0 |
| NVIDIA AI-Q Deployment ✓ NVIDIA · Official | 57 · Use with care | ★ 3.6k | 13d ago | Apache-2.0 |
Unlike doca-comch, this skill handles socket-shaped data across the host-DPU boundary, while comch provides control-plane messaging. Unlike doca-eth, it preserves a socket interface; doca-eth is the named alternative when an application can be rewritten for line-rate raw packet I/O. DOCA installation and environment preparation belong to doca-setup.
How did FollowSkills review this skill?
The documentation explicitly separates read-only from state-changing operations, requires confirmation through single-client smoke tests, records rollback paths, and discloses data-path and forwarding-endpoint risks. However, sensitive-log handling is incomplete, the instruction to retain “full unredacted” captures is risky, and privilege and external-dependency boundaries partly rely on the public guide, so points are deducted.
The task structure, failure taxonomy, and escalation paths are fairly detailed. Static evidence also shows an internal inconsistency: one section expects connection-set inspection while another states that the shipped binary has no corresponding listing flags. Key commands and outputs are omitted, preventing static reproduction; the score is therefore capped below 10 and reduced further.
Triggers, audience, scenarios, and out-of-scope boundaries are clearly stated, with AF_UNIX, BlueField, and container deployment considerations. However, the skill depends on DOCA, BlueField, Linux, and external public documentation, and gives no Chinese-language or mainland-China reachability assessment, so points are deducted.
The skill uses layered SKILL.md, CAPABILITIES.md, TASKS.md, evaluation, and skill-card materials, with dependencies, limitations, routing, and maintenance contacts. Author metadata is not clear in the skill itself, version and license statements conflict across files, and the benchmark reports missing recommended sections, hierarchy issues, and unexpected files, so points are deducted.
It offers useful deployment decisions, configuration guidance, smoke testing, and diagnosis structure. However, it intentionally omits complete commands, defaults, and sample outputs, requiring the installed version’s --help and external documentation; substantial operator-side supplementation and review remain necessary, limiting the static score.
BENCHMARK.md, evaluation cases, and summarized results provide limited third-party execution evidence. The underlying dataset is unavailable and cannot be independently reproduced, while many technical claims depend on public documentation not included in the supplied material, leaving evidence coverage thin.
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 →