DOCA Flow Perf (doca_flow_perf)
Where to start: This is a tool skill for invoking
doca_flow_perf, the host-side / DPU-CPU-side DOCA Flow
performance measurement tool. Open TASKS.md and
start at ## configure to commit to
the three-axis decision (target Flow pipeline shape × traffic
class × measurement axis) and pick the JSON policy file that
expresses the workload, then ## run for the
single-iteration smoke, then ## test for
the iterative eval loop that produces a defensible
Kops/sec-class number. Open CAPABILITIES.md
when the question is what doca_flow_perf measures and what
it deliberately does not measure, how its DPDK and DOCA
backends differ behind the same JSON contract, how to
interpret the per-iteration CPU-cycle output, or how it
differs from doca-flow-tune (measurement vs. optimization)
and doca-flow-dpa-perf (host / DPU-CPU vs. DPA-offloaded
path). If DOCA is not installed, route to
doca-setup first; if the
target measurement is the DPA-offloaded path, route to
doca-flow-dpa-perf
instead; if the goal is to optimize an already-deployed Flow
pipeline rather than measure a synthetic one, route to
doca-flow-tune — flow-perf
is a synthetic-driver microbenchmark, not a tuner of a live
Flow application.
Example questions this skill answers well
- "I want a defensible host-side baseline number for how
many
doca-flow rules per second a single BlueField-3 can
insert for a 5-tuple match-and-hairpin workload. Which
policy JSON do I start from, how do I make the result
reproducible, and what do I have to capture alongside the
number for it to be defensible?" — class-shaped flow-perf
baseline question; the agent walks the configs/ library,
the JSON contract, and the four-tuple capture rule.
- "What is the difference between
doca-flow-perf,
doca-flow-dpa-perf, and doca-flow-tune? They all
mention doca-flow and perf in their names — when do I
reach for each?" — measurement-vs-optimization plus
host-vs-DPA-path; the agent surfaces the boundaries.
- "My policy JSON looks like the example, but the reported
Kops/sec is dramatically lower than the published numbers I
see in NVIDIA's release notes. What variables do I have to
control before I can trust the comparison?" — methodology
question; the agent walks the controllable axes (number of
workers, queue depth, burst size, fixed-vs-incremented match
fields, DPDK vs DOCA backend, BlueField mode, driver /
firmware).
- "I have a workload that does not match any of the shipped
policy JSONs in
configs/. How do I author a new policy
JSON, what is the JSON schema in broad strokes, and what
changes when I switch a match field from mode: fixed to
mode: increase?" — JSON authoring question; the agent
walks the shipped configs as exemplars and refuses to
invent schema fields not present in the source tree.
- "What does the tool actually NOT measure? I am trying to
understand whether a flow-perf number tells me anything
about end-to-end traffic latency or just about the
rule-programming control-plane rate." — methodology
perimeter question; the agent draws a hard line: this tool
measures rule install / delete (control-plane) rate plus
optional query rate, NOT dataplane latency, NOT dataplane
throughput, NOT end-to-end application performance.
- "I see two backends — DPDK and DOCA — behind the same
JSON. When do I pick which, and what does the choice mean
for the result I report?" — backend choice question; the
agent walks the DPDK-backend vs. DOCA-backend trade-off and
insists the operator REPORT which one they used.
Audience
Experienced AI agents and platform / network engineers who
are comfortable with the doca-flow programming model and
the DPDK control-plane, who want a defensible number for
the host-side / DPU-CPU-side Flow rule-install / rule-delete
rate. Readers are expected to know that the published
numbers in NVIDIA release notes are run with very specific
preconditions (specific DOCA version, specific BlueField
firmware, specific traffic class) and that any number they
produce locally must explicitly state those preconditions.
This skill is NOT for:
- operators who want to optimize an already-deployed
doca-flow application — that is
doca-flow-tune;
- operators measuring the DPA-offloaded Flow path — that is
doca-flow-dpa-perf;
- operators measuring end-to-end dataplane throughput or
latency — that is the application's responsibility,
layered on
doca-flow;
- contributors authoring or modifying the tool itself.
Language scope
User interaction with doca_flow_perf is via:
- The shipped binary's command-line flags (documented by
doca_flow_perf --help and the public DOCA Flow Perf
guide on docs.nvidia.com).
- A JSON policy file describing the pipeline (ports, pipes,
matchers, actions, forwarding). The shipped
configs/
directory contains canned policies for the most common
traffic classes; new policies are authored by copying and
editing one of those.
- The tool's per-iteration output (CPU cycles per iteration,
number-processed, number-failed; reported via the tool's
stdout — the exact format is the public guide and the
binary's runtime output, NOT this skill's invention).
The skill itself is Markdown. There is no programmatic API on
top of doca_flow_perf; consumers of its results read its
stdout / captured logs.
When to load this skill
Load doca-flow-perf when ANY of the following is true:
- the user mentions
doca_flow_perf, doca-flow-perf, the
configs/ JSON library, or asks for a "host-side flow
rules per second" number;
- the user wants to baseline an underlying Flow path (not
optimize a live application);
- the user is comparing host-side / DPU-CPU-side Flow
performance across DOCA releases, BlueField generations,
or firmware versions;
- the user wants to design a new traffic class JSON and
needs to know which canned
configs/ JSON to start from
and which fields they can change.
Co-load this skill with:
doca-flow (the
underlying library; flow-perf programs the same
matchers / actions / pipes the library exposes);
doca-flow-tune (the
measurement-vs-optimization distinction is the most
common confusion);
doca-flow-dpa-perf
(the host-vs-DPA-path distinction is the second most
common confusion);
doca-version (the
four-way version match every reported flow-perf number
must carry);
doca-debug and
doca-setup for the
env-side debug ladder.
Do NOT load this skill when the user wants to optimize a live
Flow application (route to
doca-flow-tune) or measure
the DPA-offloaded path (route to
doca-flow-dpa-perf).
What this skill provides
Three companion files in this directory, each owning a
different question shape:
SKILL.md — this file. Audience, scope,
loading order, related skills. Routes everything else.
CAPABILITIES.md — what
doca_flow_perf is, what it measures, what it
deliberately doesn't measure, the DPDK-vs-DOCA backend
duality, the JSON contract surface, the per-iteration
output interpretation, version compatibility (versioned
with doca-flow and doca-version), the layered error
taxonomy, observability, and the safety policy overlay.
TASKS.md — the procedural verbs (configure,
run, test, debug, etc.) plus a doca_flow_perf-
specific command appendix and the agent-side use
workflow that consumes the captured per-iteration output.
The combined skill teaches an AI agent to drive the
measurement-class of doca_flow_perf questions: pick a
shipped or author-new policy JSON, run the single-iteration
smoke, run the iterative eval loop, capture the four-tuple
that makes the resulting number defensible, interpret the
output, and route every adjacent question (tune the live
app, measure the DPA path, optimize the firmware) to the
right neighbouring skill.
What this skill deliberately does not ship
- End-to-end dataplane throughput or latency
measurement.
doca_flow_perf measures the
control-plane rate of programming rules, plus optional
per-entry query timing. It does NOT measure how fast
packets traverse the resulting rules in the dataplane.
That is the application's responsibility, layered on
doca-flow. The agent
must say this explicitly when the operator asks for "Flow
throughput".
- DPA-offloaded Flow path measurement. Route to
doca-flow-dpa-perf.
- Optimization of a deployed Flow application. Route to
doca-flow-tune. flow-perf
is a synthetic driver of a JSON-described pipeline, not a
tuner of a live one.
- A canonical "right answer" Kops/sec number. The agent
refuses to quote published numbers from memory as
authoritative; the published numbers live in NVIDIA's
release notes per the DOCA version and BlueField
generation, and the operator must reproduce on their own
exact preconditions before comparing.
- Invented JSON schema fields. The agent does NOT invent
policy JSON keys that are not present in the shipped
configs/ exemplars. If a key the operator wants is not
in any shipped exemplar, the agent says so and routes to
the public DOCA Flow Perf guide.
- Library-internal
doca-flow API explanations. The
underlying matchers and actions belong to
doca-flow; this skill
references them but does not duplicate the library's API
documentation.
- Cross-tool benchmarking apples-to-apples claims when
preconditions differ. Two flow-perf numbers from different
DOCA versions / BlueField generations / firmware versions
are NOT directly comparable; the agent insists on the
four-tuple capture so consumers can judge.
Loading order
When a doca_flow_perf question arrives:
- Confirm DOCA is installed and the binary plus the
configs/ JSON library are reachable — if not, route to
doca-setup;
- Confirm the underlying
doca-flow library is healthy on
the device — if not, route to
doca-flow TASKS.md ## test;
- Confirm the user wants to measure, not optimize —
if optimize, route to
doca-flow-tune;
- Confirm the target path is host / DPU-CPU, not DPA — if
DPA, route to
doca-flow-dpa-perf;
- Read
CAPABILITIES.md to commit to
the three-axis decision (pipeline shape × traffic class ×
measurement axis);
- Read
TASKS.md and walk
## configure → ## run → ## test → ## debug in that
order; do NOT start with ## run without the
## configure precondition step.
Related skills
Cross-link conventions follow the bundle's relative path
contract from tools/<X>/:
doca-flow — the
underlying library. flow-perf programs Flow pipes, entries,
matchers, and actions; the library is the source of truth
for the API surface flow-perf exercises.
doca-flow-tune — the
unified Flow tuning tool. Measurement vs. optimization
boundary lives here. Ask: "do I want a number, or do I
want to change the deployed pipeline?"
doca-flow-dpa-perf —
the DPA-offloaded Flow performance tool. Host /
DPU-CPU vs. DPA path boundary lives here. Ask: "am I
measuring the path that executes on the CPU, or the path
that executes on the DPA processor?"
doca-version — every
reported flow-perf number must come with the four-way
match (host package, kernel module, firmware, target
application's linked doca-flow version) and the BlueField
/ ConnectX generation. flow-perf overlays this rule, not
contradicts it.
doca-setup — DOCA install
posture; routing for "is the binary even here?" questions.
doca-debug — the
cross-cutting debug ladder for env-side issues (driver,
firmware, BlueField mode, kernel module).
doca-bench — a peer
benchmarking tool with a broader scope (multiple DOCA
primitives, not just Flow). flow-perf is the Flow-specific
microbenchmark; doca-bench is the broader workload
benchmark.
doca-public-knowledge-map
— routing to the public docs.nvidia.com DOCA Flow Perf
page, release notes, and forums for release-specific
published numbers and reproducibility notes.
doca-structured-tools-contract
— the agent's detect → prefer → fall back → report contract
for the structured helpers (doca-env --json,
doca-capability-snapshot, version-matrix.json)
flow-perf preconditions rely on.
doca-hardware-safety
— the canonical hardware-safety meta-policy that
CAPABILITIES.md ## Safety policy
overlays.
This skill assumes the surrounding doca-flow application is
the operator's existing source artifact; flow-perf does not
ship a sample doca-flow application of its own.
1---2license: Apache-2.03name: doca-flow-perf4description: >5 Use this skill when the user is measuring the host or DPU-CPU6 control-plane rate of a DOCA Flow pipeline with doca_flow_perf —7 picking a JSON policy from configs/, choosing the DPDK or DOCA8 backend, running the single-iteration smoke then the iterative eval9 loop, interpreting per-iteration CPU cycles and num_pushed /10 num_failed, or capturing the four-tuple (DOCA version,11 BlueField/firmware, JSON policy, worker/queue/burst config) that12 makes a Kops/sec number defensible. Trigger even when the user does13 not explicitly mention "doca-flow-perf" — typical implicit phrasings14 include "how many rules per second can my BlueField insert",15 "5-tuple hairpin rule rate", "Kops/sec for steering", "flow-perf16 number does not match release notes", "DPDK vs DOCA benchmark", or17 "rule-install variance too high". Refuse and route elsewhere for18 optimizing a live Flow app (doca-flow-tune), the DPA-offloaded path19 (doca-flow-dpa-perf), dataplane throughput or latency, or20 library-internal pipe semantics — those belong to other skills.21metadata:22 source: https://skillmd.com/skills/nvidia/doca-flow-perf23 kind: tool24compatibility: >25 Requires DOCA SDK installed at /opt/mellanox/doca on Linux (Ubuntu26 22.04/24.04 or RHEL/SLES) with a BlueField DPU or ConnectX NIC27 attached. The doca_flow_perf binary plus its configs/ JSON exemplars28 must be present (the DOCA Flow Perf install component), with the29 underlying doca-flow library healthy. Reads `pkg-config doca-flow`30 and inspects /opt/mellanox/doca/{lib,include,samples,applications}.31---3233# DOCA Flow Perf (`doca_flow_perf`)3435**Where to start:** This is a tool skill for invoking36`doca_flow_perf`, the host-side / DPU-CPU-side DOCA Flow37performance measurement tool. Open [`TASKS.md`](TASKS.md) and38start at [`## configure`](TASKS.md#configure) to commit to39the three-axis decision (target Flow pipeline shape × traffic40class × measurement axis) and pick the JSON policy file that41expresses the workload, then [`## run`](TASKS.md#run) for the42single-iteration smoke, then [`## test`](TASKS.md#test) for43the iterative eval loop that produces a defensible44Kops/sec-class number. Open [`CAPABILITIES.md`](CAPABILITIES.md)45when the question is *what `doca_flow_perf` measures and what46it deliberately does not measure*, *how its DPDK and DOCA47backends differ behind the same JSON contract*, *how to48interpret the per-iteration CPU-cycle output*, or *how it49differs from `doca-flow-tune` (measurement vs. optimization)50and `doca-flow-dpa-perf` (host / DPU-CPU vs. DPA-offloaded51path)*. If DOCA is not installed, route to52[`doca-setup`](../../doca-setup/SKILL.md) first; if the53target measurement is the DPA-offloaded path, route to54[`doca-flow-dpa-perf`](../doca-flow-dpa-perf/SKILL.md)55instead; if the goal is to optimize an already-deployed Flow56pipeline rather than measure a synthetic one, route to57[`doca-flow-tune`](../doca-flow-tune/SKILL.md) — `flow-perf`58is a synthetic-driver microbenchmark, not a tuner of a live59Flow application.6061## Example questions this skill answers well6263- *"I want a defensible host-side baseline number for how64 many `doca-flow` rules per second a single BlueField-3 can65 insert for a 5-tuple match-and-hairpin workload. Which66 policy JSON do I start from, how do I make the result67 reproducible, and what do I have to capture alongside the68 number for it to be defensible?"* — class-shaped flow-perf69 baseline question; the agent walks the `configs/` library,70 the JSON contract, and the four-tuple capture rule.71- *"What is the difference between `doca-flow-perf`,72 `doca-flow-dpa-perf`, and `doca-flow-tune`? They all73 mention `doca-flow` and `perf` in their names — when do I74 reach for each?"* — measurement-vs-optimization plus75 host-vs-DPA-path; the agent surfaces the boundaries.76- *"My policy JSON looks like the example, but the reported77 Kops/sec is dramatically lower than the published numbers I78 see in NVIDIA's release notes. What variables do I have to79 control before I can trust the comparison?"* — methodology80 question; the agent walks the controllable axes (number of81 workers, queue depth, burst size, fixed-vs-incremented match82 fields, DPDK vs DOCA backend, BlueField mode, driver /83 firmware).84- *"I have a workload that does not match any of the shipped85 policy JSONs in `configs/`. How do I author a new policy86 JSON, what is the JSON schema in broad strokes, and what87 changes when I switch a match field from `mode: fixed` to88 `mode: increase`?"* — JSON authoring question; the agent89 walks the shipped configs as exemplars and refuses to90 invent schema fields not present in the source tree.91- *"What does the tool actually NOT measure? I am trying to92 understand whether a flow-perf number tells me anything93 about end-to-end traffic latency or just about the94 rule-programming control-plane rate."* — methodology95 perimeter question; the agent draws a hard line: this tool96 measures rule install / delete (control-plane) rate plus97 optional query rate, NOT dataplane latency, NOT dataplane98 throughput, NOT end-to-end application performance.99- *"I see two backends — DPDK and DOCA — behind the same100 JSON. When do I pick which, and what does the choice mean101 for the result I report?"* — backend choice question; the102 agent walks the DPDK-backend vs. DOCA-backend trade-off and103 insists the operator REPORT which one they used.104105## Audience106107Experienced AI agents and platform / network engineers who108are comfortable with the `doca-flow` programming model and109the DPDK control-plane, who want a *defensible* number for110the host-side / DPU-CPU-side Flow rule-install / rule-delete111rate. Readers are expected to know that the published112numbers in NVIDIA release notes are run with very specific113preconditions (specific DOCA version, specific BlueField114firmware, specific traffic class) and that any number they115produce locally must explicitly state those preconditions.116117This skill is NOT for:118119- operators who want to optimize an already-deployed120 `doca-flow` application — that is121 [`doca-flow-tune`](../doca-flow-tune/SKILL.md);122- operators measuring the DPA-offloaded Flow path — that is123 [`doca-flow-dpa-perf`](../doca-flow-dpa-perf/SKILL.md);124- operators measuring end-to-end dataplane throughput or125 latency — that is the application's responsibility,126 layered on127 [`doca-flow`](../../libs/doca-flow/SKILL.md);128- contributors authoring or modifying the tool itself.129130## Language scope131132User interaction with `doca_flow_perf` is via:1331341. The shipped binary's command-line flags (documented by135 `doca_flow_perf --help` and the public DOCA Flow Perf136 guide on `docs.nvidia.com`).1372. A JSON policy file describing the pipeline (ports, pipes,138 matchers, actions, forwarding). The shipped `configs/`139 directory contains canned policies for the most common140 traffic classes; new policies are authored by copying and141 editing one of those.1423. The tool's per-iteration output (CPU cycles per iteration,143 number-processed, number-failed; reported via the tool's144 stdout — the exact format is the public guide and the145 binary's runtime output, NOT this skill's invention).146147The skill itself is Markdown. There is no programmatic API on148top of `doca_flow_perf`; consumers of its results read its149stdout / captured logs.150151## When to load this skill152153Load `doca-flow-perf` when ANY of the following is true:154155- the user mentions `doca_flow_perf`, `doca-flow-perf`, the156 `configs/` JSON library, or asks for a "host-side flow157 rules per second" number;158- the user wants to baseline an underlying Flow path (not159 optimize a live application);160- the user is comparing host-side / DPU-CPU-side Flow161 performance across DOCA releases, BlueField generations,162 or firmware versions;163- the user wants to design a new traffic class JSON and164 needs to know which canned `configs/` JSON to start from165 and which fields they can change.166167Co-load this skill with:168169- [`doca-flow`](../../libs/doca-flow/SKILL.md) (the170 underlying library; flow-perf programs the same171 matchers / actions / pipes the library exposes);172- [`doca-flow-tune`](../doca-flow-tune/SKILL.md) (the173 measurement-vs-optimization distinction is the most174 common confusion);175- [`doca-flow-dpa-perf`](../doca-flow-dpa-perf/SKILL.md)176 (the host-vs-DPA-path distinction is the second most177 common confusion);178- [`doca-version`](../../doca-version/SKILL.md) (the179 four-way version match every reported flow-perf number180 must carry);181- [`doca-debug`](../../doca-debug/SKILL.md) and182 [`doca-setup`](../../doca-setup/SKILL.md) for the183 env-side debug ladder.184185Do NOT load this skill when the user wants to optimize a live186Flow application (route to187[`doca-flow-tune`](../doca-flow-tune/SKILL.md)) or measure188the DPA-offloaded path (route to189[`doca-flow-dpa-perf`](../doca-flow-dpa-perf/SKILL.md)).190191## What this skill provides192193Three companion files in this directory, each owning a194different question shape:195196- [`SKILL.md`](SKILL.md) — this file. Audience, scope,197 loading order, related skills. Routes everything else.198- [`CAPABILITIES.md`](CAPABILITIES.md) — *what199 `doca_flow_perf` is*, what it measures, what it200 deliberately doesn't measure, the DPDK-vs-DOCA backend201 duality, the JSON contract surface, the per-iteration202 output interpretation, version compatibility (versioned203 with `doca-flow` and `doca-version`), the layered error204 taxonomy, observability, and the safety policy overlay.205- [`TASKS.md`](TASKS.md) — the procedural verbs (`configure`,206 `run`, `test`, `debug`, etc.) plus a `doca_flow_perf`-207 specific command appendix and the agent-side `use`208 workflow that consumes the captured per-iteration output.209210The combined skill teaches an AI agent to drive the211*measurement-class* of `doca_flow_perf` questions: pick a212shipped or author-new policy JSON, run the single-iteration213smoke, run the iterative eval loop, capture the four-tuple214that makes the resulting number defensible, interpret the215output, and route every adjacent question (tune the live216app, measure the DPA path, optimize the firmware) to the217right neighbouring skill.218219## What this skill deliberately does not ship220221- **End-to-end dataplane throughput or latency222 measurement.** `doca_flow_perf` measures the223 *control-plane* rate of programming rules, plus optional224 per-entry query timing. It does NOT measure how fast225 packets traverse the resulting rules in the dataplane.226 That is the application's responsibility, layered on227 [`doca-flow`](../../libs/doca-flow/SKILL.md). The agent228 must say this explicitly when the operator asks for "Flow229 throughput".230- **DPA-offloaded Flow path measurement.** Route to231 [`doca-flow-dpa-perf`](../doca-flow-dpa-perf/SKILL.md).232- **Optimization of a deployed Flow application.** Route to233 [`doca-flow-tune`](../doca-flow-tune/SKILL.md). flow-perf234 is a synthetic driver of a JSON-described pipeline, not a235 tuner of a live one.236- **A canonical "right answer" Kops/sec number.** The agent237 refuses to quote published numbers from memory as238 authoritative; the published numbers live in NVIDIA's239 release notes per the DOCA version and BlueField240 generation, and the operator must reproduce on their own241 exact preconditions before comparing.242- **Invented JSON schema fields.** The agent does NOT invent243 policy JSON keys that are not present in the shipped244 `configs/` exemplars. If a key the operator wants is not245 in any shipped exemplar, the agent says so and routes to246 the public DOCA Flow Perf guide.247- **Library-internal `doca-flow` API explanations.** The248 underlying matchers and actions belong to249 [`doca-flow`](../../libs/doca-flow/SKILL.md); this skill250 references them but does not duplicate the library's API251 documentation.252- **Cross-tool benchmarking apples-to-apples claims** when253 preconditions differ. Two flow-perf numbers from different254 DOCA versions / BlueField generations / firmware versions255 are NOT directly comparable; the agent insists on the256 four-tuple capture so consumers can judge.257258## Loading order259260When a `doca_flow_perf` question arrives:2612621. Confirm DOCA is installed and the binary plus the263 `configs/` JSON library are reachable — if not, route to264 [`doca-setup`](../../doca-setup/SKILL.md);2652. Confirm the underlying `doca-flow` library is healthy on266 the device — if not, route to267 [`doca-flow TASKS.md ## test`](../../libs/doca-flow/TASKS.md#test);2683. Confirm the user wants to *measure*, not *optimize* —269 if optimize, route to270 [`doca-flow-tune`](../doca-flow-tune/SKILL.md);2714. Confirm the target path is host / DPU-CPU, not DPA — if272 DPA, route to273 [`doca-flow-dpa-perf`](../doca-flow-dpa-perf/SKILL.md);2745. Read [`CAPABILITIES.md`](CAPABILITIES.md) to commit to275 the three-axis decision (pipeline shape × traffic class ×276 measurement axis);2776. Read [`TASKS.md`](TASKS.md) and walk278 `## configure → ## run → ## test → ## debug` in that279 order; do NOT start with `## run` without the280 `## configure` precondition step.281282## Related skills283284Cross-link conventions follow the bundle's relative path285contract from `tools/<X>/`:286287- [`doca-flow`](../../libs/doca-flow/SKILL.md) — the288 underlying library. flow-perf programs Flow pipes, entries,289 matchers, and actions; the library is the source of truth290 for the API surface flow-perf exercises.291- [`doca-flow-tune`](../doca-flow-tune/SKILL.md) — the292 unified Flow tuning tool. **Measurement vs. optimization293 boundary** lives here. Ask: "do I want a number, or do I294 want to change the deployed pipeline?"295- [`doca-flow-dpa-perf`](../doca-flow-dpa-perf/SKILL.md) —296 the DPA-offloaded Flow performance tool. **Host /297 DPU-CPU vs. DPA path boundary** lives here. Ask: "am I298 measuring the path that executes on the CPU, or the path299 that executes on the DPA processor?"300- [`doca-version`](../../doca-version/SKILL.md) — every301 reported flow-perf number must come with the four-way302 match (host package, kernel module, firmware, target303 application's linked `doca-flow` version) and the BlueField304 / ConnectX generation. flow-perf overlays this rule, not305 contradicts it.306- [`doca-setup`](../../doca-setup/SKILL.md) — DOCA install307 posture; routing for "is the binary even here?" questions.308- [`doca-debug`](../../doca-debug/SKILL.md) — the309 cross-cutting debug ladder for env-side issues (driver,310 firmware, BlueField mode, kernel module).311- [`doca-bench`](../doca-bench/SKILL.md) — a peer312 benchmarking tool with a broader scope (multiple DOCA313 primitives, not just Flow). flow-perf is the Flow-specific314 microbenchmark; doca-bench is the broader workload315 benchmark.316- [`doca-public-knowledge-map`](../../doca-public-knowledge-map/SKILL.md)317 — routing to the public `docs.nvidia.com` DOCA Flow Perf318 page, release notes, and forums for release-specific319 published numbers and reproducibility notes.320- [`doca-structured-tools-contract`](../../doca-structured-tools-contract/SKILL.md)321 — the agent's detect → prefer → fall back → report contract322 for the structured helpers (`doca-env --json`,323 `doca-capability-snapshot`, `version-matrix.json`)324 flow-perf preconditions rely on.325- [`doca-hardware-safety`](../../doca-hardware-safety/SKILL.md)326 — the canonical hardware-safety meta-policy that327 [`CAPABILITIES.md ## Safety policy`](CAPABILITIES.md#safety-policy)328 overlays.329330This skill assumes the surrounding doca-flow application is331the operator's existing source artifact; flow-perf does not332ship a sample doca-flow application of its own.