GPUNetIO RDMA Write Latency Benchmark
Guides reproducible measurement of CUDA-kernel-initiated GPUNetIO RDMA WRITE latency and tail behavior.
The skill scopes the task and discloses GPU/NIC, CUDA, DOCA, TCP, and RDMA data flows; it also addresses trusted OOB networking, handle lifetime, and routing of hardware changes through a safety policy. It lacks explicit user-confirmation gates, least-privilege guidance, rollback procedures, and detailed sensitive-output handling, while license metadata is inconsistent, so points are deducted.
The documentation provides prerequisites, build/run sequencing, layered error taxonomy, and failure routing, making the happy path plausible. However, it depends on companion files, real DOCA hardware, and a prepared installation; some output and iteration claims are not reproducibly specified, and static review cannot validate key paths, so the score is capped and reduced.
Triggers, audience, non-fit boundaries, and the GPUNetIO/GPI/CPU distinction are comparatively clear. Chinese-language support is not demonstrated, and some guidance depends on external DOCA documentation and specialized hardware; reachability from mainland-China networks is not established, so points are deducted.
The information architecture, loading order, task routing, limitation disclosures, and maintenance contacts are useful. Author, stable version, changelog, and explicit update ownership are incomplete; the SKILLCARD contains TBD fields, and benchmark paths, repository paths, and license declarations are inconsistent, so points are deducted.
The skill can materially guide GPU-initiated RDMA WRITE latency measurement, statistic selection, topology qualification, and troubleshooting. It is a thin loader without scripts, parsers, or representative expected outputs, requires user-side compilation on specialized hardware, and leaves substantial result preparation to the user, so points are deducted.
The files cite common.h, meson.build, source layout, and an evaluation report, providing some traceability. The key companion source files are not included in the supplied evidence, the evaluation dataset is unavailable, and the report is not independently reproducible CI/test evidence, so only a low score is justified.
- No commands were executed; key build, runtime, distribution, and failure paths were not independently reproduced.
- SKILL.md states Apache-2.0, while the SKILLCARD states Apache 2.0 AND CC-BY-4.0, and the supplied license metadata is NOASSERTION; this should be reconciled before publication.
- The benchmark report lacks its source dataset and per-task records, so PASS and percentage results cannot be treated as independent verification.
- The documentation requests full unredacted stdout without clearly specifying how host, network, or environment-sensitive data should be handled.
- Reachability of external DOCA documentation in the target user environment should be confirmed, and Chinese-language triggers and guidance should be added.
What does this skill do, and when should you use it?
This skill is for developers and performance engineers evaluating RDMA WRITE latency initiated from a GPU kernel through doca-gpunetio. It covers the `gpunetio_ib_write_lat` client/server ping-pong benchmark, including build and run guidance, GPU-NIC pairing, and environment checks. It explains the half-iter, full-iter, and CUDA-side usec columns and the use of median, p99, and jitter. It is useful for measured infrastructure baselines, but it does not represent the latency of an entire real-time application pipeline.
Guides the agent to check the DOCA SDK, CUDA Toolkit, nvidia_peermem, GPU-NIC topology, and pkg-config dependencies; build the C/CUDA client and server with Meson from /opt/mellanox/doca/tools/gpunetio_ib_write_lat; run the two-sided RDMA WRITE ping-pong benchmark; interpret half-iter, full-iter, and CUDA-side usec output; characterize median, p99, and jitter; and compare GPUNetIO with GPI and CPU-initiated perftest.
- A real-time control-loop developer validates whether GPU-initiated WRITE tail latency meets a deadline between two GPU and ConnectX-equipped hosts.
- A platform operator reruns the benchmark after changing NUMA placement, GPU PCIe topology, IB device selection, GID index, or NIC firmware.
- A performance engineer documents median, p99, jitter, software versions, hardware, and deployment conditions for regression analysis.
- A team chooses between the GPUNetIO and GPI programming surfaces or the CPU-initiated `perftest` path.
- A CUDA developer evaluates the latency and GPU-side overhead trade-off of batching multiple work requests.
What are this skill's strengths and limitations?
- Narrowly focused on CUDA-kernel-initiated RDMA WRITE latency through doca-gpunetio.
- Covers build, smoke testing, evaluation, output semantics, and tail-latency reporting.
- Encourages recording hardware, software versions, and deployment state for reproducibility.
- Clearly distinguishes GPUNetIO, GPI, and CPU-initiated `perftest` paths.
- Requires a GPU-NIC pair on each endpoint plus DOCA, CUDA, `nvidia_peermem`, and InfiniBand-capable hardware.
- The skill is agent guidance rather than a scripts, parser, bindings, or samples bundle.
- It intentionally omits fixed flags and expected latency values; the installed binary and source must be checked.
- The benchmark measures GPUNetIO WR latency, not the full application pipeline or end-to-end control-loop deadline.
- The supplied material includes no platform test results, test suite, or benchmark numbers.
How do you install this skill?
Install the skill from the NVIDIA/skills collection with: npx skills add nvidia/skills --skill doca-gpunetio-ib-write-lat --yes. The host must also meet the stated Linux, DOCA SDK, CUDA Toolkit, GPU, InfiniBand-capable ConnectX or BlueField RNIC, and loaded nvidia_peermem requirements. The source does not provide a standalone installer or complete DOCA installation command.
How do you use this skill?
After installation, ask the agent something like: “Measure GPUNetIO GPU-init RDMA WRITE latency between two H100 + ConnectX hosts and report median, p99, and jitter separately.” The agent should verify the GPU-NIC and DOCA/CUDA prerequisites, then follow the referenced TASKS.md configure, build, run, and test workflows. Use the installed binary’s --help and main.c registration as the authority for exact flags.
How does this skill compare with similar options?
GPUNetIO is the runtime surface benchmarked here. GPI is an alternative programming surface for the same physical operation, but doca/tools/ does not ship a GPI ib_write_lat benchmark binary. Upstream perftest ib_write_lat measures CPU-initiated WRITE latency. The related doca-gpunetio-ib-write-bw skill covers bandwidth rather than latency.