version: "1.0.0" name: crank-benchmarks description: Build, review, and troubleshoot crank benchmark commands for aspnet/Benchmarks and custom scenarios. Use when configuring multi-agent benchmark runs (application/db/load), selecting --scenario and --profile combinations, adding diagnostics (counters/traces/dumps), pinning .NET versions/commits, and saving comparable benchmark outputs with --json.

Crank Benchmarks

Overview

Use this skill to produce reliable crank commands, explain profile/scenario wiring, and capture comparable results across repeated runs. Default to framework-band targeting (--application.framework) and persisted JSON output unless the user asks for deeper pinning.

Workflow

1. Identify benchmark intent (scenario, scale, target machines, diagnostics).
2. Resolve required configs and profile mappings before assembling command arguments, preferring predefined profile groups when available.
3. For regression investigations, validate version availability from package feeds before running:

• Query the relevant flat-container index (for example .../flat2/microsoft.netcore.app.runtime.linux-x64/index.json) to list actually published versions.
• Cross-check benchmarked versions vs published versions to detect missing intermediate builds.
• For each candidate intermediate version, fetch the package .nuspec and read <repository ... commit=\"...\"> to map the version to a concrete dotnet/dotnet commit.
• Use these commit anchors to reduce and prioritize culprit ranges even when some benchmark runs are missing.

1. Build a single canonical command line with:

• --config files
• one --scenario
• required --profile values for each role
• optional diagnostics and optional version-pinning switches (for regression reproduction)
• optional git source pinning with --application.source.branchOrCommit
• --json <output.json> for persisted results

1. Validate command completeness against scenario roles (application, optional db, load).
2. Produce follow-up compare commands when prior JSON files are available.

Installing and updating crank

Crank is a .NET global tool and needs the .NET SDK and git. Only the controller (crank) is needed to run benchmarks against existing agents (including the gold machines through relay profiles); Microsoft.Crank.Agent is only for hosting your own agent.

# install
dotnet tool install -g Microsoft.Crank.Controller --version "0.2.0-*"
# update to the latest 0.2.0 prerelease
dotnet tool update -g Microsoft.Crank.Controller --version "0.2.0-*"
# verify
crank --help

For a local agent, do the same with Microsoft.Crank.Agent and run crank-agent. Check the installed version with dotnet tool list -g before troubleshooting unexpected option errors; update if it is outdated. Docs: https://github.com/dotnet/crank/blob/main/docs/getting_started.md

Quick-start example

Minimal APIs JSON (TechEmpower) on the Linux gold machines:

crank \
  --config https://raw.githubusercontent.com/aspnet/Benchmarks/main/src/BenchmarksApps/TechEmpower/Minimal/minimal.benchmarks.yml \
  --scenario json \
  --profile aspnet-gold-lin \
  --application.framework net11.0 \
  --application.options.collectCounters true \
  --json minimal-json.json

The Platform variant uses --config .../scenarios/platform.benchmarks.yml instead; see references/techempower-scenarios.md.

Defaults and guardrails

Follow these defaults unless the user asks otherwise:

• Keep one explicit --scenario.
• Prefer one deterministic profile group (for example aspnet-gold-lin or aspnet-gold-win, defined in scenarios/aspnet.profiles.standard.yml) before role-specific profile tuples.
• When the user is not on the benchmark network (or says relay), use the -relay twin of the profile together with --relay (bare --relay authenticates via az login); see references/techempower-scenarios.md.
• Include exactly the profiles needed to satisfy scenario roles and machine placement.
• Add --json <file> for every run meant to be compared later.
• Use --application.framework net10.0|net11.0 as the normal version selector; pin SDK/runtime/aspnetcore versions only when reproducing a .NET regression.
• When pinning versions for regressions, prefer versions confirmed in package feed indexes rather than assuming all CI build labels were published.
• If benchmark history skips versions, fetch .nuspec metadata for missing versions to map to source commits before narrowing culprit changes.
• When using a git-based benchmark source, use --application.source.branchOrCommit <branch_name> or --application.source.branchOrCommit #<commit_hash> to target specific code.
• Add --application.collectDependencies true only for CI transitive dependency analysis.
• Add build/azure.profile.yml only when Azure machines/relay/auth-specific profiles are required.
• Prefer additive diagnostics (--*.collectCounters, traces, dumps) over changing scenario behavior.
• If result comparison is requested, include --compare <prior.json> arguments or a follow-up crank compare step.

Reference loading

Read references based on task:

• references/api_reference.md for command argument semantics and repeatable command templates.
• references/profiling-and-traces.md for CPU profiles, dotnet-trace/perf traces, counters and dumps.
• references/techempower-scenarios.md for the TechEmpower Minimal APIs and Platform plaintext/json/fortunes scenarios, their config files, and CI trend names.

Output expectations

• Return a final crank command that is directly runnable.
• Include short rationale for each --profile selected.
• Call out when optional version pins are intentionally omitted or included.
• Include output filenames for JSON persistence and comparison readiness.