What problem does it solve?
Measuring the latency of RDMA WRITE work requests posted from CUDA kernels through doca-gpunetio requires building a source-shipped client/server benchmark, validating GPU-NIC PCIe pairing, and correctly interpreting half-iter, full-iter, and CUDA-side timing columns. This Skill guides agents through that entire workflow so latency numbers for real-time control-loop decisions are reproducible and defensible.
Core Features & Use Cases
- Build and run guidance: Walks through meson-based compilation of the gpunetio_ib_write_lat client and server pair against the installed DOCA SDK, plus smoke-before-bulk execution.
- Latency characterization: Teaches median vs p99 vs jitter statistic selection, the latency-vs-batching trade-off intrinsic to GPU-initiated RDMA, and sound measurement practices.
- Surface selection and debugging: Helps choose between GPUNetIO, GPI, and CPU-initiated perftest surfaces, and provides a layered error taxonomy from config syntax through RDMA connection failures.
- Use Case: A performance engineer validating whether an H100 + ConnectX-7 pair meets a real-time control-loop deadline uses this Skill to build the benchmark, verify GPU-NIC pairing, and capture a median/p99/jitter distribution with the full version tuple.
Quick Start
Ask your agent to measure GPU-kernel-initiated RDMA WRITE latency between two hosts using the doca-gpunetio ib_write_lat benchmark and report median and p99 latency.
