Perf Core

Do not skip steps. Do not optimize without profiling first.

Universal Tools

Quick benchmarking with hyperfine

# Single command
hyperfine 'my-program --input data.json'

# Compare two implementations
hyperfine 'my-program-v1 input.txt' 'my-program-v2 input.txt'

# Warmup runs and minimum iterations
hyperfine --warmup 3 --min-runs 10 'my-program'

Reading Flame Graphs

• Width = time spent in that function (wider = hotter)
• Height = call stack depth (taller = deeper call chain)
• Look for wide plateaus -- these are hot functions consuming the most time
• Ignore narrow towers -- deep but fast, not the bottleneck
• Use differential flame graphs to compare before/after optimization
• Colors are arbitrary in most tools -- width is the only metric that matters

Performance Regression Testing

• Run benchmarks in CI on every pull request
• Compare results against a baseline from the default branch
• Alert on regressions exceeding a defined threshold (e.g., >5%)
• Store benchmark results as CI artifacts for historical comparison
• Use statistical analysis (multiple runs, confidence intervals) to avoid false positives from noise

Anti-Patterns

Language-Specific Skills

After establishing methodology with this skill, load the appropriate language-specific skill for concrete tooling: