DOCA Perftest
When to load this skill
Load this skill when the user wants to — or the agent needs to — run
doca_perftest on a host with DOCA installed. Concretely:
- Measuring RDMA bandwidth, message rate, or latency between two hosts
(point-to-point) with a verb such as write / read / send.
- Running the same benchmark across many nodes — JSON scenario or
--traffic_pattern (pair-wise, all-to-all), orchestrated over
OpenMPI / passwordless SSH.
- Comparing the IBV and DOCA Verbs backends (
-r ibv vs -r dv) or
driving the GPUNetIO engine (--engine gpunetio).
- Interpreting a run: client-vs-server number drift, "device busy",
inconsistent latency across runs.
Do not load this skill for DOCA install (doca-setup), the doca-bench harness,
Flow pipeline tuning (doca-flow-tune), or application-level
throughput.
Examples
- "Measure RDMA write bandwidth between two ConnectX-7 hosts with 64 KB messages." Point-to-point: run the server, then the client with
-n.
- "I need latency across 16 nodes, all-to-all." Multi-node: JSON scenario or
--traffic_pattern, with OpenMPI.
- "Compare the IBV and DOCA Verbs drivers on the same device." Same command,
-r ibv vs -r dv.
Instructions
Follow the loading order below. Open TASKS.md for the commands and CAPABILITIES.md for what the tool can measure and what it does not.
Where to start: DOCA Perftest is the DOCA 3.5 replacement for legacy ib_write_bw / ib_send_lat plus the orchestration scripts used to run them on many nodes. Start at TASKS.md ## point-to-point for a first run. Use ## multi-node for scenarios and ## debug when a run fails or the numbers look wrong.
If DOCA is not installed, route to doca-setup/SKILL.md. Client and server must run the same DOCA Perftest version; check with doca-version/SKILL.md before comparing numbers across hosts. If the question is about GPU-initiated RDMA in the user's own code, route to doca-gpunetio/SKILL.md.
Example questions this skill answers well
- "Run a point-to-point RDMA write bandwidth test." Use
-d <dev> -n <server> -v write -m bw -s <size> -D <sec>; the client coordinates the server side over SSH (requires SSH key exchange set up beforehand).
- "Benchmark 16 nodes at once and get one report." Use a JSON scenario or
--traffic_pattern; the report aggregates per device, node and test.
- "Latency differs between runs." Check the latency mode, the core and NUMA placement, and whether both sides use the same message size and verb.
Loading order
- This file for scope and routing.
TASKS.md for the exact command for the job you are doing.
CAPABILITIES.md when you need to know whether a verb, driver, memory type or traffic pattern is supported.
Do not open every companion file.
Scope boundaries
- Installs and driver packages:
doca-setup/SKILL.md.
- Version matching across client, server and DOCA release:
doca-version/SKILL.md.
- Any change to NIC or eswitch state, firmware, or host kernel settings (hugepages, IOMMU, PCIe rebind): follow
doca-hardware-safety/SKILL.md first.
- Failure diagnosis:
doca-debug/SKILL.md for the layer ladder, then the TASKS.md ## debug anchor here.
- Upstream CPU-initiated
ib_write_bw: this tool replaces it on DOCA 3.5; do not mix the two in one comparison without saying so.
1---2license: Apache-2.03name: doca-perftest4description: >5 Use this skill when the user is running DOCA Perftest (doca_perftest), the6 RDMA benchmark that measures bandwidth, message rate and latency for RDMA7 verbs, point-to-point or across a cluster, with an optional DOCA Verbs8 backend (-r dv) or GPUNetIO engine (--engine gpunetio). Trigger even when the9 user does not say "DOCA Perftest" — implicit phrasings include "how do I10 benchmark RDMA write bandwidth with DOCA", "run ib_write_bw but across 811 nodes", "perftest says the device is busy", "my latency numbers differ between12 client and server", "launch the server over SSH", or "run an all-to-all RDMA13 test from a JSON scenario". Refuse and route elsewhere for DOCA install14 (doca-setup), DOCA Bench (doca-bench), and application-level throughput.15metadata:16 source: https://skillmd.com/skills/nvidia/doca-perftest17 author: NVIDIA <doca-feedback@nvidia.com>18 kind: tool19 doca_version: "3.5"20 tags:21 - doca22 - rdma23 - benchmark24 - perftest25compatibility: >26 Requires DOCA 3.5 or later on Linux with a ConnectX or BlueField RDMA device,27 libibverbs (DOCA-OFED or upstream rdma-core), and OpenMPI for multi-node28 modes. The DOCA Verbs backend needs the doca-verbs package; the GPUNetIO29 engine needs doca-gpunetio, CUDA and a Volta+ GPU. Run benchmarks only on a30 trusted, non-shared fabric.31---3233# DOCA Perftest3435## When to load this skill3637Load this skill when the user wants to — or the agent needs to — run38`doca_perftest` on a host with DOCA installed. Concretely:3940- Measuring RDMA bandwidth, message rate, or latency between two hosts41 (point-to-point) with a verb such as write / read / send.42- Running the same benchmark across many nodes — JSON scenario or43 `--traffic_pattern` (pair-wise, all-to-all), orchestrated over44 OpenMPI / passwordless SSH.45- Comparing the IBV and DOCA Verbs backends (`-r ibv` vs `-r dv`) or46 driving the GPUNetIO engine (`--engine gpunetio`).47- Interpreting a run: client-vs-server number drift, "device busy",48 inconsistent latency across runs.4950Do **not** load this skill for DOCA install (`doca-setup`), the doca-bench harness,51Flow pipeline tuning (`doca-flow-tune`), or application-level52throughput.5354## Examples5556- "Measure RDMA write bandwidth between two ConnectX-7 hosts with 64 KB messages." Point-to-point: run the server, then the client with `-n`.57- "I need latency across 16 nodes, all-to-all." Multi-node: JSON scenario or `--traffic_pattern`, with OpenMPI.58- "Compare the IBV and DOCA Verbs drivers on the same device." Same command, `-r ibv` vs `-r dv`.5960## Instructions6162Follow the loading order below. Open `TASKS.md` for the commands and `CAPABILITIES.md` for what the tool can measure and what it does not.6364**Where to start:** DOCA Perftest is the DOCA 3.5 replacement for legacy `ib_write_bw` / `ib_send_lat` plus the orchestration scripts used to run them on many nodes. Start at `TASKS.md ## point-to-point` for a first run. Use `## multi-node` for scenarios and `## debug` when a run fails or the numbers look wrong.6566If DOCA is not installed, route to `doca-setup/SKILL.md`. Client and server must run the same DOCA Perftest version; check with `doca-version/SKILL.md` before comparing numbers across hosts. If the question is about GPU-initiated RDMA in the user's own code, route to `doca-gpunetio/SKILL.md`.6768## Example questions this skill answers well6970- **"Run a point-to-point RDMA write bandwidth test."** Use `-d <dev> -n <server> -v write -m bw -s <size> -D <sec>`; the client coordinates the server side over SSH (requires SSH key exchange set up beforehand).71- **"Benchmark 16 nodes at once and get one report."** Use a JSON scenario or `--traffic_pattern`; the report aggregates per device, node and test.72- **"Latency differs between runs."** Check the latency mode, the core and NUMA placement, and whether both sides use the same message size and verb.7374## Loading order75761. This file for scope and routing.772. `TASKS.md` for the exact command for the job you are doing.783. `CAPABILITIES.md` when you need to know whether a verb, driver, memory type or traffic pattern is supported.7980Do not open every companion file.8182## Scope boundaries8384- Installs and driver packages: `doca-setup/SKILL.md`.85- Version matching across client, server and DOCA release: `doca-version/SKILL.md`.86- Any change to NIC or eswitch state, firmware, or host kernel settings (hugepages, IOMMU, PCIe rebind): follow `doca-hardware-safety/SKILL.md` first.87- Failure diagnosis: `doca-debug/SKILL.md` for the layer ladder, then the `TASKS.md ## debug` anchor here.88- Upstream CPU-initiated `ib_write_bw`: this tool replaces it on DOCA 3.5; do not mix the two in one comparison without saying so.