doca-gpunetio-ib-write-lat - Measure GPUNetIO GPU-initiated RDMA WRITE latency
Builds and runs the GPUNetIO ib_write_lat client-server benchmark to measure GPU-kernel-initiated RDMA WRITE latency, median, p99, and jitter.
Tags
Updated: 2026-09-29Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Build the benchmark binaries
- Check GPU-NIC pairing
- Run client-server latency tests
- Read latency output columns
- Characterize median, p99, and jitter
- Compare GPUNetIO, GPI, and CPU-init paths
Inputs
- DOCA source tree
- Installed DOCA SDK
- CUDA Toolkit
- GPU-NIC pair
- Client and server hosts
- Benchmark run parameters
Outputs
- Built client and server binaries
- Latency measurements
- Median, p99, and jitter results
- Build and invocation records
- Benchmark failure diagnostics
Requirements
- Linux host
- DOCA SDK at /opt/mellanox/doca
- InfiniBand-capable ConnectX or BlueField RNIC
- NVIDIA GPU with CUDA Toolkit
- Loaded nvidia_peermem module
- GPU-NIC pair on each host
- Common PCIe or NVLink fabric
- Matching DOCA and CUDA dependencies
- DOCA pkg-config modules