skills.rest
NVIDIA
NVIDIAOFFICIAL
Explore all skills by this author →
WebsiteGitHub

doca-bench-extension

Author, build, and debug custom doca-bench plug-in extensions for uncovered workload classes.

3.2k|370|Updated Feb 25, 2026
Search Tags:#benchmarking#shared library#CUDA#doca#bluefield#gpunetio#doca-bench
One-click install
npx skills add https://github.com/NVIDIA/skills --skill doca-bench-extension
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: doca-bench-extension
Source: https://github.com/NVIDIA/skills/tree/main/skills/doca-bench-extension
Command: npx skills add https://github.com/NVIDIA/skills --skill doca-bench-extension

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

When none of doca-bench's built-in workload modes can measure your workload, you must author a custom extension — a versioned shared library with DOCA_EXPERIMENTAL-marked C entry points — and getting the build, registration, loading, and debugging right is error-prone. This Skill guides agents through that entire extension lifecycle using the shipped doca_bench_cuda reference exemplar.

Core Features & Use Cases

  • Extension decision and design: Confirms whether a workload genuinely requires an extension versus a built-in doca-bench mode, then walks the API surface shape (entry points, settings structs, accounting structs, stop_flag lifetime contract) using the shipped doca_bench_cuda reference.
  • Build and version management: Guides Meson-based builds of versioned shared libraries whose soversion matches the running DOCA release, including the DOCA_EXPERIMENTAL rebuild-on-upgrade rule.
  • Layered debugging: Diagnoses failures across seven layers — build, load, registration, runtime call, and version mismatches — with a smoke-before-bulk testing discipline.
  • Use Case: A performance engineer needs to benchmark a CUDA GPUNetIO RX/TX kernel that no built-in doca-bench mode covers; the Skill walks them through copying the reference exemplar, adapting its entry points, building the .so, and running a no-op smoke before scaling up.

Quick Start

Ask your agent to help you benchmark a workload that no built-in doca-bench mode covers by authoring a custom extension based on the shipped doca_bench_cuda reference.

Frequently Asked Questions about doca-bench-extension

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I benchmark a workload that no built-in doca-bench mode covers?▼

Author a custom doca-bench extension: a versioned shared library exposing DOCA_EXPERIMENTAL C entry points that doca-bench loads at runtime. Start by copying the shipped doca_bench_cuda reference under /opt/mellanox/doca/tools/bench_extension/ and adapting its API surface shape.

How do I benchmark a CUDA GPUNetIO RX/TX kernel with doca-bench?▼

Use the shipped doca_bench_cuda reference extension, which drives GPUNetIO RX/TX queues from CUDA kernels with per-workload settings structs and a stop_flag lifetime contract. If your workload matches its nop/eth-recv/eth-send/eth-bidir kernels, you can use it directly without writing a new extension.

Why can't doca-bench find or load my custom extension .so?▼

Load failures usually come from an soversion mismatch with the running DOCA release, the library not being on the dynamic linker's search path, or an architecture/ABI mismatch. Check the SONAME with readelf -d, run ldd on the library, and confirm it was built against the same DOCA release's headers.

Do I need to rebuild my doca-bench extension after a DOCA upgrade?▼

Yes. The extension API is marked DOCA_EXPERIMENTAL, so the surface may change across releases without a deprecation window, and the library's soversion must match the running DOCA so_version. Rebuilding against the new release's headers and re-running the smoke tests is the rule on every DOCA upgrade.

When should I use a built-in doca-bench mode instead of an extension?▼

Always check the built-in mode inventory first — extensions are the exit ramp only when no built-in mode measures your workload class. Extensions add real cost: a build toolchain, version churn across DOCA releases, and the experimental API stability contract.

Why does my GPU extension kernel hang during benchmarking?▼

The most common cause is the stop_flag never being set by the parent, so the CUDA kernel runs forever and recovery requires a GPU reset. Every long-running extension kernel must respect the parent-set stop signal, and OOB reset access is a non-negotiable precondition for GPU-side extensions.

Related Skills

Explore More Software Engineering Skills→
air-gappedair-gapped
5

nvidia-datacenter-bringup

Automate NVIDIA datacenter GPU bring-up on Ubuntu 24.04 with health validation.

Official
Advanced
thetanilthetanil
1

benchmark

Benchmark HBF endpoints with Apache Bench and store results in SQLite.

Community
Advanced
gqf2008gqf2008

modal-serverless-gpu

Deploy ML workloads on serverless GPU infrastructure with auto-scaling.

Community
Advanced
MadAppGangMadAppGang
7

audit

Audit codebases for security vulnerabilities, code smells, and compliance issues.

Official
Advanced