Skill de Benchmark de Performance Linux
jc1122/perf-benchmark-skill
Benchmark e diagnóstico de performance com rubrica multi-dimensional
Ver no GitHubSkill para agentes de código que perfila cargas de trabalho Python e C em Linux, avaliando performance através de rubrica de 7 dimensões (escalabilidade algorítmica, estabilidade de tempo, eficiência CPU, comportamento de cache, predição de branch, perfil de memória e ASM). Integra-se com ferramentas nativas como Valgrind e perf para diagnóstico profundo de gargalos.
Instalação
npx convez-ai jc1122/perf-benchmark-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Análise de escalabilidade algorítmica em Python/C com múltiplos tamanhos de entrada
- Diagnóstico de ineficiências de cache, branch prediction e memória
- Validação de otimizações de performance em pipelines CI/CD
- Detecção de regressões de performance com baseline ledger
- Perfilamento hotspot nativo via perf record/report
Veredito

Ferramenta rigorosa e bem instrumentada para profissionais que precisam otimizar performance em nível de infraestrutura Linux. Exige conhecimento sólido de conceitos de performance (cache, branch prediction, ASM) e é mais adequada para contextos de CI/CD e otimização sistemática do que uso casual.
README
A tradução para o português deste README ainda está na fila.
perf-benchmark-skill
Linux performance benchmarking skill for coding agents that support Skills. It profiles Python and C workloads, scores a 7-dimension rubric, and keeps algorithmic issues ahead of cache, branch, and ASM tuning.
For repo-agnostic use, pass an explicit --target or --binary.
Installation
npx skills add <skill-source>/perf-benchmark-skill
<skill-source> means the installable source or repository path that hosts this skill.
Scope
- Algorithmic scaling, wall-time stability, CPU efficiency, cache behavior, branch prediction, memory profile, and optional ASM review
- Linux-only, using
/proc,/sys, Valgrind, andperf SKILL.mdis the detailed agent-facing workflow and CLI reference
Pytest benchmark autodiscovery is a convenience for Python repos. For
non-pytest entrypoints, use --target or --binary.
Multi-size explicit targets must include {SIZE}.
Full Algorithmic Scaling scoring requires deep or asm because allocation churn comes from massif.
Usage
python scripts/perf_benchmark_pipeline.py \
--root /path/to/repo \
--out-dir /tmp/bench \
--tier deep \
--target "python -m benchmark_entrypoint {SIZE}" \
--sizes 10000,100000 \
--source-prefix path/to/source/ \
--perf-record \
--max-cv 5.0 \
--findings-out /tmp/perf-findings.json \
--baseline-ledger /tmp/perf-ledger.jsonl
--perf-record is opt-in native hotspot sampling via perf record and
perf report. Use it when perf is available and you want flat sampled
hotspots in addition to the rubric.
--max-cv (default 5.0) sets the coefficient-of-variation noise gate:
timing-derived dimensions exceeding this threshold are scored N/A (noise).
--findings-out writes shared-schema PERF findings JSON (one per FAIL/WARN
dimension, signal: "PERF"). --baseline-ledger maintains an append-only
JSONL run history with vs-last and vs-best regression checks; can be used
alongside --baseline for point-in-time comparison.
Outputs
benchmark_report.md: scorecard, findings, prescriptionsbenchmark_summary.json: machine-readable scores and regression dataperf_findings.json: shared-schema PERF findings (when--findings-outset)baseline_ledger.jsonl: append-only run history (when--baseline-ledgerset)tier1/totier4/: raw profiler artifacts by depth
More Detail
See SKILL.md for the full workflow, tier behavior, agent guidance, and reference links.
Related Skills
perf-optimization (v0.1.0)
The perf-optimization/ directory contains a companion
skill that consumes perf-benchmark findings and applies an iterative
measure -> change -> re-measure ratchet to systematically resolve diagnosed
bottlenecks. It selects the highest-impact candidate per iteration, makes one
bounded change, re-runs profiling under identical conditions, and records
accepted wins in an append-only ledger. Algorithmic scaling failures gate all
constant-factor work.
See perf-optimization/SKILL.md for workflow details and verification requirements.
License
MIT