Doca Ngauge
NVIDIAOfficialCreated by NVIDIA
- Skill
- Tool integration
- Software development
- Apache-2.0
- Medium risk
AutoGineer summary
Written by AutoGineer from the skill’s own files, not by the author
The doca-ngauge skill provides guidance for using the DOCA ngauge tool to probe NIC hardware counters on supported NVIDIA devices, including BlueField-3 and ConnectX-7. It enables users to collect performance data in HDF5 format, visualize metrics, and debug network performance on a single node. The skill covers YAML-based configuration, execution workflows, and the interpretation of counter data, while maintaining strict boundaries against hardware state changes or raw packet I/O operations.
What it does
- Probes NIC hardware counters for network performance analysis.
- Stores collected performance telemetry in HDF5 format.
- Supports YAML-based configuration for sampling parameters and counter selection.
- Provides methods for visualizing and parsing collected HDF5 data.
- Interprets CLI progress indicators and counter summary statistics.
- Diagnoses common runtime issues such as fwctl driver conflicts.
When to use it
- When you need to measure network performance on a single node using hardware counters.
- When configuring sampling periods and counter IDs via YAML files.
- When you need to plot or analyze previously collected ngauge HDF5 output.
- When troubleshooting performance issues that require deep NIC hardware telemetry.
- When resolving conflicts involving the fwctl driver.
Advantages
- Enables granular network performance monitoring via hardware-level probes.
- Standardizes data collection using the HDF5 format for consistent analysis.
- Provides a structured approach to debugging performance bottlenecks.
- Integrates with existing DOCA toolsets for streamlined operations.
Usage examples
- “Probe NIC hardware counters on a BlueField-3 and save HDF5 data.”
- The agent will guide you through creating a YAML configuration file with the appropriate device address and counter IDs, then show you how to execute the ngauge command.
- “Plot the collected counter data from an ngauge output file.”
- The agent will direct you to use the sample plugins provided in the ngauge installation to parse and visualize your HDF5 data.
- “ngauge cannot start because fwctl is in use.”
- The agent will help identify conflicting services using the fwctl driver and advise on how to resolve the dependency.
Install
Download the skill as a ZIP:
https://autogineer.com/agent-skills/nvidia/doca-ngauge/downloadUnzip it into the cross-client skills folder of your project:
mkdir -p .agents/skills && unzip doca-ngauge.zip -d .agents/skills
Documentation by NVIDIA
DOCA ngauge
Scope boundaries
These requests are out of scope: say so, name the skill that owns the request, and stop.
raw Ethernet packet send/receive at line rate:
doca-eth/SKILL.md.host-DPU control-plane messaging (Comch):
doca-comch/SKILL.md.DOCA install and bring-up:
doca-setup/SKILL.md.which DOCA version is installed, version compatibility:
doca-version/SKILL.md.NIC or BlueField mode and firmware changes:
doca-hardware-safety/SKILL.md.general DOCA orientation and documentation lookup:
doca-public-knowledge-map/SKILL.md.
Examples
"Probe NIC hardware counters on a BlueField-3 and save HDF5 data." — this is in scope. Configure the YAML file and run ngauge; do not attempt to change hardware or firmware state here.
"Plot the collected counter data from an ngauge output file." — use the sample plugins under
/usr/share/doc/ngauge/examples/plugins; dependencies may need separate installation."ngauge cannot start because fwctl is in use." — check for other tools or services that also use fwctl; route install and driver conflict resolution to
doca-setup.
Quick facts
The authoritative source for the binary name, flags, paths, and
configuration schema is the public DOCA ngauge guide at
https://docs.nvidia.com/doca/archive/3-5-0/doca-ngauge/index.html.
Confirm details with the installed tool's --help and the sample
configuration files before quoting them.
ngauge analyzes, visualizes, and debugs network performance on a single node by probing NIC hardware counters and storing collected data in HDF5 format with relevant metadata.
Supported hardware: NVIDIA BlueField-3, NVIDIA ConnectX-7, and above; firmware version xx.43.1000 or higher.
ngauge relies on the fwctl driver and cannot run simultaneously with other tools or services that also use fwctl.
All configurations are defined in an input YAML file. Sample configuration files are under
/usr/share/doc/ngauge/examples/settings.The device is specified by its PCIe address, for example
device: "0000:03:00.0".Output path and file prefix are mandatory. The output file is saved as
/path/to/output/directory/<prefix>_<DATE>_<TIME>.h5; the exact file name is printed after each run.The
silentoption suppresses CLI progress indications when set totrue(defaultfalse).Runtime parameters include
mode(repetitiveorsingle) andperiod_us(sampling period in microseconds; decimal or scientific notation accepted, e.g.1e2= 100 μs).Counter definitions require
id(data ID); optional fields aredesc,unit,accumulating,normalizer(time, a non-zero number, or a configured data ID with anid/prefix),cutoff_min, andcutoff_max.All supported performance counters are listed in the public guide's "Supported Data IDs" section.
Run ngauge with:
ngauge <configuration YAML file>To end a run early, use Ctrl+C (SIGINT). This is a normal and supported way to stop; all results collected up to that point are saved as usual.
During a run, progress bars for each counter are displayed. Color coding: blue for low values, red for high values, gradient for intermediate values, solid gray for no changes. Data is updated once per second, showing mean, standard deviation (σ), minimum, and maximum observed during the last second.
To reference the most recent results when the output directory is
/tmp(the default), use:$(ls -1 /tmp/ngauge_data_*.h5 | tail -n1)
Instructions
Follow Loading order below. Open the tasks companion for configure / run / parse / debug. Open the capabilities companion for what ngauge measures and reports. Do not open every companion file.
Where to start: This is a tool skill for invoking DOCA ngauge —
the single-node network performance analysis, visualization, and
debugging tool that probes NIC hardware counters and stores HDF5 data.
Open the tasks companion and start at ## configure for the YAML
configuration decision, then ## run for the execution flow. Open the
capabilities companion when the question is what counters can be
measured, what metadata is stored, and what the progress display
implies. If the user has not installed ngauge or fwctl, route to
doca-setup/SKILL.md first. If the user is asking about
hardware/firmware state changes, route to
doca-hardware-safety/SKILL.md. If the user is asking about version
compatibility, route to doca-version/SKILL.md.
How to use this skill. Read this file, then search the companion
file for the ## <name> heading the question needs and read that whole
section. Answer completely from it, quoting only its symbols and
commands. Do not browse docs; use the entry points in
doca-public-knowledge-map. Give environment checks as commands for
the user's DOCA host. For out-of-scope requests, name the skill that
owns them and stop.
Example questions this skill answers well
The CLASSES of ngauge questions this skill is built to answer, each with one worked example. The class is the load-bearing piece; the worked example is one instance.
"How do I configure ngauge to sample specific NIC hardware counters?" — worked example: "I need to measure RX bytes port 0 and TX bytes port 0 on a BlueField-3 at a 100 μs period, save to /tmp with prefix ngauge_data_." Answered by the YAML configuration model in the capabilities companion + the tasks companion step 2.
"How do I run ngauge and stop it early without losing data?" — worked example: "I started ngauge with a YAML config, want to stop after 10 seconds and still get HDF5 output." Answered by the execution flow and Ctrl+C behavior in the tasks companion.
"How do I plot the collected HDF5 data?" — worked example: "I have /tmp/ngauge_data_2026-09-01_12-00-00.h5 and want to plot counter 0x1020000100000000." Answered by the parse/plot workflow in the tasks companion using
simple_plot.pyorsimple_text_plot.py."What do the progress bar colors mean?" — worked example: "During a run, some bars are red, some blue, some gray; what does that indicate?" Answered by the progress display semantics in the capabilities companion.
"Why can't ngauge run alongside another tool?" — worked example: "ngauge fails to start because fwctl is in use by another service." Answered by the fwctl dependency and conflict section in the capabilities companion + route to
doca-setup."What hardware and firmware are required?" — worked example: "Will ngauge work on ConnectX-6 or BlueField-2?" Answered by the prerequisites in the capabilities companion.
When to load this skill
Load this skill when the user is — or the agent needs to — drive DOCA ngauge on a real host or BlueField Arm with ngauge installed. Concretely:
Probing NIC hardware counters on BlueField-3 or ConnectX-7 and above.
Configuring the ngauge YAML file (device, output, params, counters).
Running ngauge and interpreting the CLI progress bars and summary.
Parsing or plotting the HDF5 output with the sample plugins.
Diagnosing fwctl driver conflicts or missing prerequisites.
Cross-checking counter data against expected network performance.
Do not load this skill for general DOCA orientation, raw packet
I/O, the comch control plane, DOCA install, version questions, or
hardware/firmware state changes. For those, route to
doca-public-knowledge-map/SKILL.md,
doca-eth/SKILL.md,
doca-comch/SKILL.md,
doca-setup/SKILL.md,
doca-version/SKILL.md, or
doca-hardware-safety/SKILL.md.
What this skill provides
This is a thin loader. Substantive material lives in two companion files:
the capabilities companion — what ngauge measures and reports: supported hardware and firmware, the fwctl dependency and conflict, the YAML configuration model (device, output, params, counters), counter field semantics (
id,desc,unit,accumulating,normalizer,cutoff_min,cutoff_max), HDF5 output format, progress display semantics, sample plugins, and the safety posture that ngauge itself is read-only probing while any step that changes hardware or firmware state routes todoca-hardware-safety.the tasks companion — step-by-step workflows for the in-scope task verbs:
configure(YAML setup),run(execution and Ctrl+C),parse/plot(simple_plot.py and simple_text_plot.py),debug(fwctl conflict, missing dependencies), plus aCommand appendixthat honors the bundle'sdoca-structured-tools-contract/SKILL.mdpreamble.
The skill assumes a host or BlueField where ngauge and fwctl are already installed and the operator has whatever privileges the public DOCA ngauge guide requires for the chosen device.
What this skill deliberately does not ship
This skill is agent guidance, not a samples or scripts bundle. To keep the boundary clean, it deliberately does not contain — and pull requests should not add:
Verbatim counter ID inventories, flag lists, or output column names. The public DOCA ngauge guide on
docs.nvidia.comand the installed--helpon the user's version are the joint source of truth; copying them here pins the skill to one release and silently rots when ngauge evolves. The skill routes the agent at those sources instead.Pre-baked example output. Output is install-, version-, and device-specific. A captured example will mislead an operator on a different platform or state.
Wrappers, parsers, or scripts in any language that consume ngauge output. The output format is documented; users who want to script against it should read the live guide and write the parser against their installed version.
A
samples/orreference/subtree. This is a thin loader for a documented DOCA artifact; substantive material lives on the public page and in--help.
Loading order
Read this skill first to confirm the user's question is in scope (the user wants to probe NIC hardware counters, collect HDF5 data, or plot ngauge output, not learn the eth / comch / RDMA APIs).
For what ngauge measures, the YAML configuration model, the read-only vs state-changing split, version availability, the fwctl conflict surface, observability, and safety posture, see the capabilities companion.
For the documented invocations and the configure → run → parse/plot workflow —
configure,run,parse,debug, plus theCommand appendix— see the tasks companion.
Related skills
doca-setup/SKILL.md— env preparation, install verification, fwctl driver installation and conflict resolution, and the I have no install yet path. This skill assumes its preconditions are satisfied.doca-version/SKILL.md— canonical DOCA version-handling rules. The## Version compatibilitysection in the capabilities companion is a concise overlay that redirects here for the body.doca-hardware-safety/SKILL.md— any step that changes hardware or firmware state routes here. ngauge itself is read-only probing, but if a workflow includes firmware updates or device state changes, use this skill.doca-public-knowledge-map/SKILL.md— routing to the public DOCA ngauge guide and the rest of the public DOCA documentation set. The canonical URL is reached viadoca-public-knowledge-map ## DOCA tools(doca-public-knowledge-map/SKILL.md#doca-tools).doca-structured-tools-contract/SKILL.md— the bundle's detect → prefer → fall back → report contract for structured helper tools. The Command appendix in the tasks companion honors this contract.doca-debug/SKILL.md— the cross-cutting debug ladder. ngauge slots in at the runtime layer as a single-node telemetry tool whose fwctl conflict surface escalates to the cross-cutting ladder when the cause is below DOCA.doca-eth/SKILL.md— the line-rate raw packet I/O surface. ngauge is for hardware counter telemetry, not packet I/O; use doca-eth when the application needs to operate on packets directly.doca-comch/SKILL.md— the host ↔ DPU control-plane primitive. ngauge is single-node performance telemetry, not a control-plane channel; pair them only when coordinating a measurement across host and DPU.
Files
- ATTRIBUTION.mdAdded by AutoGineer891 B
- BENCHMARK.md10.9 KiB
- CAPABILITIES.md12.2 KiB
- LICENSELicence file added by AutoGineer11.2 KiB
- SKILL.md13.7 KiB
- SKILLCARD.yaml2.4 KiB
- TASKS.md13.7 KiB
- evals/evals.json3.0 KiB
- skill-card.md4.8 KiB
- skill.oms.sig5.1 KiB
CAPABILITIES.md
DOCA ngauge — Capabilities
Where to start: ngauge is a single DOCA tool for analyzing,
visualizing, and debugging network performance on a single node by
probing NIC hardware counters and storing collected data in HDF5
format. Pick the pattern first, then drill into the H2 that owns the
substance. For the how of executing each pattern, jump to
TASKS.md. For the public guide that is the authoritative
source for binary names, flags, paths, and the configuration schema,
see ## Public-source pointer.
This file is loaded by SKILL.md. It documents what
ngauge measures and reports, which configuration axes are
documented, what versions and prerequisites it ships with, the
layered error and observability surfaces, and the safety policy that
gates any state-changing step. For step-by-step invocations and the
configure → run → parse/plot workflow, see TASKS.md.
Pattern overview
Every ngauge question this skill teaches resolves into one of SIX patterns. The patterns are CLASSES — they apply across every DOCA install that ships ngauge, not just one platform.
| ngauge pattern | Class shape | Where the substance lives |
|---|---|---|
| 1. Configure a counter sampling run | Choose the device PCIe address, output path and prefix, runtime parameters (mode, period_us), and the counter list in the input YAML file | ## Capabilities and modes configuration model + TASKS.md ## configure |
| 2. Run and stop a measurement | Execute ngauge <configuration YAML file>; use Ctrl+C (SIGINT) to end early and still save all collected data | TASKS.md ## run + ## Safety policy read-only posture |
| 3. Plot or parse HDF5 output | Use the sample plugins simple_plot.py or simple_text_plot.py against the generated .h5 file and one or more counter IDs | ## Capabilities and modes sample plugins + TASKS.md ## parse |
| 4. Diagnose fwctl conflict or missing prerequisites | Map the symptom (ngauge cannot start, fwctl in use, unsupported hardware or firmware) to the right layer of the error taxonomy before any state-changing intervention | ## Error taxonomy + TASKS.md ## debug |
| 5. Interpret progress display and measurement summary | Read per-counter progress bars, color coding, and the once-per-second mean / standard deviation / minimum / maximum updates | ## Observability + ## Capabilities and modes progress display |
| 6. Route state-changing, install, or version questions | Send hardware or firmware state changes to doca-hardware-safety, installs to doca-setup, version questions to doca-version | ## Safety policy + SKILL.md ## Related skills |
Two cross-cutting rules that apply to every pattern above:
ngauge is read-only probing. It probes NIC hardware counters and stores HDF5 data; it does not change hardware or firmware state. Any step that changes hardware or firmware state routes through
doca-hardware-safety/SKILL.mdbefore execution.ngauge relies on the fwctl driver and cannot run simultaneously with other tools or services that also use fwctl. The agent must check for fwctl conflicts before any run, and must route driver installation or conflict resolution to
doca-setup/SKILL.md.
Capabilities and modes
The DOCA ngauge tool is a documented DOCA artifact that ships on
installs which include it. The exact installed location, binary name,
packaging shape, and per-platform availability live in the public DOCA
ngauge guide reachable via
## Public-source pointer
and on the installed --help; the skill deliberately does not pin
them — see
SKILL.md ## What this skill deliberately does not ship.
Use case — what ngauge is for. ngauge analyzes, visualizes, and debugs network performance on a single node. It probes NIC hardware counters and stores the collected data in HDF5 format, along with relevant metadata, for subsequent processing. The tool also provides graphical progress updates and a measurement summary directly on the CLI, offering real-time insights into the measurement process.
Configuration model — the load-bearing concept. Every ngauge run commits to a point in this space; omitting any axis produces an ambiguous answer the agent must flag back to the user.
| Axis | What it picks | Why the agent must name it |
|---|---|---|
| Device | The PCIe address of the NIC to probe, for example device: "0000:03:00.0" | Without a device, ngauge has no target; the address must be the operator's actual device, not a placeholder |
| Output | The output directory path and file prefix, both mandatory; the file is saved as /path/to/output/directory/<prefix>_<DATE>_<TIME>.h5 | The exact file name is printed after each run; omitting path or prefix makes the run invalid |
| Runtime parameters | mode (repetitive or single) and period_us (sampling period in microseconds; decimal or scientific notation accepted, e.g. 1e2 = 100 μs) | These control how long and how often ngauge samples; the agent must state them explicitly |
| Counters | A list of counter definitions; id (data ID) is mandatory, while desc, unit, accumulating, normalizer, cutoff_min, and cutoff_max are optional | The counter IDs determine what is measured; all supported performance counters are listed in the public guide's "Supported Data IDs" section |
Counter field semantics. Each counter entry requires an id
(data ID). Optional fields are:
desc— description of the counter.unit— unit type.accumulating— whether the counter accumulates values.normalizer— if present, must be eithertime, a non-zero number, or one of the configured Data IDs with anid/prefix.cutoff_min— data below this value will not be recorded.cutoff_max— data above this value will not be recorded.
Output and metadata. The output file is saved in HDF5 format with
relevant metadata. The exact file name is printed after each run. If
the silent option is set to true, progress indications on the
command line are suppressed (default false).
Progress display. During a run, progress bars for each counter are displayed. Color coding indicates value levels relative to other values of the same counter: blue for low values, red for high values, intermediate gradient colors for values between low and high, and solid gray bars for counters whose values did not change during the run. Data is updated once per second, showing the mean, standard deviation (σ), minimum, and maximum values observed during the last second.
Sample plugins. A sample plugin named simple-plot is provided
and installed under /usr/share/doc/ngauge/examples/plugins. It
demonstrates how to open the output HDF5 file generated by ngauge and
plot the data. An alternative plugin, simple_text_plot.py, generates
text-based plots directly in the terminal; its resolution is
significantly lower, but it supports zooming into specific regions and
is useful when graphical output is unavailable or network connectivity
to the server is limited. The sample plugins are provided as examples
and are not integral parts of the ngauge tool. Dependencies such as
NumPy, H5py, Matplotlib, plotext, and others may need to be installed
separately to run these plugins.
Version compatibility
The canonical version-handling rules live in
doca-version/SKILL.md; this section is a concise overlay that
redirects there for the body.
Supported hardware: NVIDIA BlueField-3, NVIDIA ConnectX-7, and above.
Required firmware version: xx.43.1000 or higher.
fwctl driver must be installed on the host. On hosts running Linux kernel version 6.15 and above, fwctl is available by default unless it was explicitly disabled during kernel configuration.
ngauge is installed with
sudo apt-get install ngaugeorsudo dnf install ngaugeon x86 or Arm 64 hosts. On the DPU, the ngauge package is pre-installed, so the above step is not necessary.This skill documents the DOCA 3.5.0 ngauge guide. For questions about other DOCA versions or version-specific behavior, route to
doca-version/SKILL.md.
Error taxonomy
Map the symptom to the right layer before any state-changing intervention.
| Symptom | Layer | Route |
|---|---|---|
| ngauge cannot start because fwctl is in use | fwctl driver conflict | Check for other tools or services that also use fwctl; route driver conflict resolution to doca-setup/SKILL.md |
| fwctl driver is missing | Prerequisite gap | Install fwctl via the OS package manager; on Linux kernel 6.15+ it is available by default unless explicitly disabled |
| Unsupported hardware or firmware below xx.43.1000 | Version / hardware gap | Confirm BlueField-3 or ConnectX-7 or above and firmware xx.43.1000 or higher; route version questions to doca-version/SKILL.md |
| Output path or prefix missing | Configuration error | Both output path and file prefix are mandatory in the YAML file |
| Invalid counter ID or normalizer | Configuration error | Counter id is mandatory; normalizer, if present, must be time, a non-zero number, or a configured Data ID with an id/ prefix |
| Plugin fails to plot | Missing plugin dependencies | Dependencies such as NumPy, H5py, Matplotlib, plotext, and others may need to be installed separately |
| Run ended before full dataset collected | Normal operation | Ctrl+C (SIGINT) is a normal and supported way to end the run; all results collected up to that point are saved as usual |
Observability
ngauge exposes the following observability surfaces:
CLI progress bars for each counter, updated once per second, showing mean, standard deviation (σ), minimum, and maximum values observed during the last second.
Color coding on progress bars: blue for low values, red for high values, gradient for intermediate values, solid gray for no changes.
Measurement summary directly on the CLI, offering real-time insights into the measurement process.
Exact output file name printed after each run.
HDF5 output file with relevant metadata for subsequent processing.
Sample plotting plugins under
/usr/share/doc/ngauge/examples/pluginsfor graphical or text-based plots. The graphical plugin includes a zoom-in function; the lighter area around samples represents the worst-case measurement error.Silent mode via the
silentoption to suppress CLI progress indications when set totrue.
To reference the most recent results when the output directory is
/tmp (the default), use the expression shown below:
$(ls -1 /tmp/ngauge_data_*.h5 | tail -n1)Safety policy
ngauge is read-only probing. It does not change hardware or firmware state. Any step that changes hardware or firmware state routes through
doca-hardware-safety/SKILL.mdbefore execution.fwctl conflict is a hard gate. ngauge cannot run simultaneously with other tools or services that also use fwctl. The agent must check for conflicts before any run and must not recommend running ngauge alongside another fwctl user.
Ctrl+C is a supported stop path. Ending a run early with SIGINT is normal and supported; all results collected up to that point are saved as usual. The agent must not treat an early stop as a failure.
Install and version questions route elsewhere. Installs route to
doca-setup/SKILL.md; version questions route todoca-version/SKILL.md. This skill assumes ngauge and fwctl are already installed and the operator has the privileges the public guide requires for the chosen device.
Public-source pointer
The authoritative source for the binary name, flags, paths, and
configuration schema is the public DOCA ngauge guide at
https://docs.nvidia.com/doca/archive/3-5-0/doca-ngauge/index.html.
Confirm details with the installed tool's --help and the sample
configuration files under /usr/share/doc/ngauge/examples/settings
before quoting them. The skill deliberately does not pin verbatim
counter ID inventories, flag lists, or output column names; those live
in the public guide and the installed version.
Requirements and compatibility
Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta
“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.
Needs: Needs Node.js
The author says: Requires a host with ngauge installed (sudo apt-get install ngauge or sudo dnf install ngauge on x86 or Arm 64 hosts; pre-installed on the DPU) and a BlueField-3 or ConnectX-7 or above device with firmware version xx.43.1000 or higher. The fwctl driver must be installed on the host; on Linux kernel 6.15 and above it is available by default unless explicitly disabled during kernel configuration. ngauge cannot run simultaneously with other tools or services that also use fwctl.
Security
What our scanner noticed· Medium risk
- Large base64-encoded blobs
A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.
Attribution and licence
- Created by
- NVIDIA
- Publisher
- NVIDIA (Official)
- Source
- NVIDIA/skills · skills/doca-ngauge · commit fce3cca
- Licence
- Apache License 2.0 (redistribution allowed)
- LICENSE (added by AutoGineer from LICENSE-APACHE)
- ATTRIBUTION.md (written by AutoGineer)
- Synced
AutoGineer is not affiliated with or endorsed by NVIDIA or NVIDIA.
Download ZIP
The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download doca-ngauge.zip