Performance Optimization

Skill: performance-optimization

Measure First — the gate every optimization passes through

Profile before optimizing — intuition about hot spots is measurably wrong more often than right:

• Correctness before speed -- make it work, then make it fast
• <4% of code causes >50% of runtime (Knuth 1971) -- find the hot spot before touching anything
• >50% of optimizations produce negligible or negative results -- measurement prevents wasted effort

Optimization proceeds only from a measurement — unmeasured "optimizations" are edits with a coin-flip chance of helping.

Scope Limitations

This skill covers single-threaded, single-process code tuning for general-purpose computing.

Not covered (need specialized guidance):

• Concurrency: Lock contention often dominates; profile thread states, not just CPU
• Distributed systems: Network latency ~10,000x memory; optimize RPC/serialization first
• Real-time systems: Need worst-case latency, not average; caching adds variance
• Embedded/constrained: Memory/power budgets require different tradeoffs

The Simplicity-Performance Relationship

Simpler code usually runs faster. Fewer special cases = less code to check; deep modules = more work per call with fewer layer crossings; complicated code does extraneous or redundant work.

Primary Workflow: 7-Step Decision Tree

Each step is a gate — skipping one forfeits the evidence the next step depends on.

1. Is the program correct and complete?
   NO  -> Make it correct first. STOP optimization.
   YES -> Continue

2. Have you measured to find the actual bottleneck?
   NO  -> Profile/measure first. Do NOT guess.
   YES -> Continue

3. Can requirements be relaxed?
   YES -> Relax requirements. Done.
   NO  -> Continue

4. Can design/architecture solve it? (Stage 2: Fundamental Fixes)
   YES -> Fix design. Done.
   NO  -> Continue

5. Can algorithm/data structure solve it?
   YES -> Change algorithm. Done.
   NO  -> Continue

6. Can compiler flags help? (40-59% improvement possible)
   YES -> Enable optimizations. Measure.
   NO  -> Continue

7. Is it in the <4% that causes >50% of runtime?
   NO  -> Do NOT optimize this code. Find actual hot spot.
   YES -> PROCEED with code tuning (see below)

Step 2 Detail: Measurement

What counts as valid measurement:

• Actual profiling data (timing, call counts, memory usage)
• Multiple runs to account for variance
• Specific hotspot identification, not just "it's slow"

Identify WHICH dimension: throughput, latency, memory, or CPU. Different problems need different solutions.

Step 4 Detail: Fundamental Fixes (APOSD Stage 2)

Before code-level changes, check for architectural fixes:

• Add a cache? Eliminate repeated expensive computation
• Better algorithm? e.g., balanced tree vs. list, hash map vs. linear search
• Bypass layers? e.g., kernel bypass for networking, direct buffer access

If a fundamental fix exists, implement it with standard design techniques. If not, continue down the tree.

Step 4 Extended: Critical Path Redesign (APOSD Stage 3)

When no fundamental fix is available, redesign the critical path:

1. Ask: What is the smallest amount of code for the common case?
2. Disregard existing code structure entirely
3. Ignore special cases in current code -- consider only data needed for critical path
4. Define "the ideal" -- simplest and fastest code assuming complete redesign freedom
5. Design the rest of the class around these critical paths

Consolidation techniques:

Code Tuning Procedure (ordered — each step's evidence feeds the next)

Only reached after completing the 7-step decision tree.

1. Save working version (cannot revert without backup)
2. Make ONE change (multiple changes = unmeasurable)
3. Measure improvement (same workload, before/after)
4. Keep if faster, revert if not (no "close enough")
5. Repeat

Technique Priority by Category

Logic:

1. Stop testing when answer known (break, short-circuit)
2. Order tests by frequency (most common first)
3. Substitute table lookups for complex logic
4. Use lazy evaluation

Loops:

1. Unswitch (move invariant tests outside)
2. Jam/fuse loops operating on same range
3. Put busiest loop on inside
4. Minimize work inside loops
5. Use sentinel values for search loops
6. Unroll ONLY if measured (can be -27% in Python!)

Data:

1. Use integers instead of floating-point when possible
2. Use fewest array dimensions
3. Cache frequently computed values
4. Precompute results where practical

Expressions:

1. Initialize at compile time
2. Exploit algebraic identities
3. Use strength reduction (multiplication -> addition)
4. Eliminate common subexpressions

After Making Changes

Checklist and code examples: Read(${CLAUDE_SKILL_DIR}/checklists.md)

Re-measure before keeping any change. Keep only if: significant speedup (with data), OR simpler AND at least as fast. Otherwise back it out.

Red Flags

Quick Reference

Checker

Checklist: Read(${CLAUDE_SKILL_DIR}/checklists.md)

Output Format:

Severity: VIOLATION (clear anti-pattern), WARNING (needs measurement), PASS (no issues)

Chain