domains / families / devops / doca-gpunetio-ib-write-lat

doca-gpunetio-ib-write-lat

separate

Use this skill when the user characterizes RDMA WRITE round-trip time posted from a CUDA kernel via GPUNetIO — compiling `gpunetio_ib_write_lat` under `doca/tools/`, verifying GPU–NIC affinity, and interpreting the half-iteration / full-iteration / nanosecond columns for median / p99 / tail-jitter analysis of a real-time control loop. Also load it for implied asks like "p99 WRITE time on H100 + ConnectX", "tail delay versus batching", or "device-originated versus upstream ib_write_lat". Refuse and route to the proper owner for throughput runs (doca-gpunetio-ib-write-bw), the GPI channel API (doca-gpi), GPUNetIO API work (doca-gpunetio), or SDK setup.

When to use

• RDMA WRITE round-trip time characterization
• Compiling gpunetio_ib_write_lat under doca/tools/
• Verifying GPU–NIC affinity
• Interpreting half-iteration / full-iteration / nanosecond columns
• Median / p99 / tail-jitter analysis
• Real-time control loop latency analysis
• p99 WRITE time on H100 + ConnectX
• Tail delay versus batching analysis

When not to use

• Throughput runs
• GPI channel API
• GPUNetIO API work
• SDK setup

Skill page

Source