Performance Regression Analysis with Flamegraphs

When to Use

• Optimizing performance-critical code
• Detecting performance regressions after changes
• Establishing performance baselines for reference
• Investigating performance issues or slow code paths
• Before creating commits with performance-sensitive changes
• When user says "check performance", "analyze flamegraph", "detect regressions", etc.

Instructions

Follow these steps to analyze performance and detect regressions:

Step 1: Generate Current Flamegraph

Run the automated benchmark script to collect current performance data:

./run.fish run-examples-flamegraph-fold --benchmark

What this does:

• Runs an 8-second continuous workload stress test
• Samples at 999Hz for high precision
• Tests the rendering pipeline with realistic load
• Generates flamegraph data in: tui/flamegraph-benchmark.perf-folded

Implementation details:

• The benchmark script is in script-lib.fish
• Uses an automated testing script that stress tests the rendering pipeline
• Simulates real-world usage patterns

Step 2: Compare with Baseline

Compare the newly generated flamegraph with the baseline:

Baseline file:

tui/flamegraph-benchmark-baseline.perf-folded

Current file:

tui/flamegraph-benchmark.perf-folded

The baseline file contains:

• Performance snapshot of the "current best" performance state
• Typically saved when performance is optimal
• Committed to git for historical reference

Step 3: Analyze Differences

Compare the two flamegraph files to identify regressions or improvements:

Key metrics to analyze:

1. Hot path changes

   • Which functions appear more/less frequently?
   • New hot paths that weren't in baseline?

2. Sample count changes

   • Increased samples = function taking more time
   • Decreased samples = optimization working!

3. Call stack depth changes

   • Deeper stacks might indicate unnecessary abstraction
   • Shallower stacks might indicate inlining working

4. New allocations or I/O

   • Look for memory allocation hot paths
   • Unexpected I/O operations

Step 4: Prepare Regression Report

Create a comprehensive report analyzing the performance changes:

Report structure:

# Performance Regression Analysis

## Summary
[Overall performance verdict: regression, improvement, or neutral]

## Hot Path Changes
- Function X: 1500 → 2200 samples (+47%) ⚠️ REGRESSION
- Function Y: 800 → 600 samples (-25%) ✅ IMPROVEMENT
- Function Z: NEW in current (300 samples) 🔍 INVESTIGATE

## Top 5 Most Expensive Functions

### Baseline
1. render_loop: 3500 samples
2. paint_buffer: 2100 samples
3. diff_algorithm: 1800 samples
...

### Current
1. render_loop: 3600 samples (+3%)
2. paint_buffer: 2500 samples (+19%) ⚠️
3. diff_algorithm: 1700 samples (-6%) ✅
...

## Regressions Detected
[List of functions with significant increases]

## Improvements Detected
[List of functions with significant decreases]

## Recommendations
[What should be investigated or optimized]

Step 5: Present to User

Present the regression report to the user with:

• ✅ Clear summary (regression, improvement, or neutral)
• 📊 Key metrics with percentage changes
• ⚠️ Highlighted regressions that need attention
• 🎯 Specific recommendations for optimization
• 📈 Overall performance trend

Optional: Update Baseline

When to update the baseline:

Only update when you've achieved a new "best" performance state:

1. After successful optimization work
2. All tests pass
3. Behavior is correct
4. Ready to lock in this performance as the new reference

How to update:

# Replace baseline with current
cp tui/flamegraph-benchmark.perf-folded tui/flamegraph-benchmark-baseline.perf-folded

# Commit the new baseline
git add tui/flamegraph-benchmark-baseline.perf-folded
git commit -m "perf: Update performance baseline after optimization"

See baseline-management.md for detailed guidance on when and how to update baselines.

Understanding Flamegraph Format

The .perf-folded files contain stack traces with sample counts:

main;render_loop;paint_buffer;draw_cell 45
main;render_loop;diff_algorithm;compare 30

Format:

• Semicolon-separated call stack (deepest function last)
• Space + sample count at end
• More samples = more time spent in that stack

Performance Optimization Workflow

1. Make code change
   ↓
2. Run: ./run.fish run-examples-flamegraph-fold --benchmark
   ↓
3. Analyze flamegraph vs baseline
   ↓
4. ┌─ Performance improved?
  │  ├─ YES → Update baseline, commit
  │  └─ NO  → Investigate regressions, optimize
  └→ Repeat

Additional Performance Tools

For more granular performance analysis, consider:

cargo bench

Run benchmarks for specific functions:

cargo bench

When to use:

• Micro-benchmarks for specific functions
• Tests marked with #[bench]
• Precise timing measurements

cargo flamegraph

Generate visual flamegraph SVG:

cargo flamegraph

When to use:

• Visual analysis of call stacks
• Identifying hot paths visually
• Sharing performance analysis

Requirements:

• flamegraph crate installed
• Profiling symbols enabled

Manual Profiling

For deep investigation:

# Profile with perf
perf record -F 999 --call-graph dwarf ./target/release/app

# Generate flamegraph
perf script | stackcollapse-perf.pl | flamegraph.pl > flame.svg

Common Performance Issues to Look For

When analyzing flamegraphs, watch for:

1. Allocations in Hot Paths

render_loop;Vec::push;alloc::grow 500 samples  ⚠️

Problem: Allocating in tight loops Fix: Pre-allocate or use capacity hints

2. Excessive Cloning

process_data;String::clone 300 samples  ⚠️

Problem: Unnecessary data copies Fix: Use references or Cow<str>

3. Deep Call Stacks

a;b;c;d;e;f;g;h;i;j;k;l;m 50 samples  ⚠️

Problem: Too much abstraction or recursion Fix: Flatten, inline, or optimize

4. I/O in Critical Paths

render_loop;write;syscall 200 samples  ⚠️

Problem: Blocking I/O in rendering Fix: Buffer or defer I/O

Reporting Results

After performance analysis:

• ✅ No regressions → "Performance analysis complete: no regressions detected!"
• ⚠️ Regressions found → Provide detailed report with function names and percentages
• 🎯 Improvements found → Celebrate and document what worked!
• 📊 Mixed results → Explain trade-offs and recommendations

Supporting Files in This Skill

This skill includes additional reference material:

• baseline-management.md - Comprehensive guide on when and how to update performance baselines: when to update (after optimization, architectural changes, dependency updates, accepting trade-offs), when NOT to update (regressions, still debugging, experimental code, flaky results), step-by-step update process, baseline update checklist, reading flamegraph differences, example workflows, and common mistakes. Read this when:
  • Deciding whether to update the baseline → "When to Update" section
  • Performance improved and want to lock it in → Update workflow
  • Unsure if baseline update is appropriate → Checklist
  • Need to understand flamegraph diff signals → "Reading Flamegraph Differences"
  • Avoiding common mistakes → "Common Mistakes" section

Related Skills

• check-code-quality - Run before performance analysis to ensure correctness
• write-documentation - Document performance characteristics

Related Commands

• /check-regression - Explicitly invokes this skill

Related Agents

• perf-checker - Agent that delegates to this skill

Additional Resources

• Flamegraph format: tui/*.perf-folded files
• Benchmark script: script-lib.fish
• Visual flamegraphs: Use flamegraph.pl to generate SVGs