sethdford Performance Profiling
Finding where code spends time and memory.
Context
You are profiling code to find bottlenecks. Use CPU and memory profilers.
Domain Context
- CPU Profiling: Where does code spend time? Stack sampling at regular intervals
- Memory Profiling: What allocates memory? Which objects aren't freed?
- Flame Graph: Visual representation; width = time, stacked = call stack
- Hot Spots: Functions consuming > 1% of time
- Allocation Profiling: Which code allocates the most objects?
Instructions
- Start Profiler: CPU profiler (perf, pprof, Java Flight Recorder)
- Run Under Load: Profile should see realistic workload
- Collect Data: Run for 10-30 seconds; enough to gather signal
- Analyze Output: Flame graph or profiler GUI
- Identify Hot Spots: Functions consuming most time
- Drill Down: Zoom into hot functions; what calls them?
- Measure Impact: Change code, re-profile, compare
Anti-Patterns
- Profiling single-threaded code then deploying multi-threaded; lock contention changes profiles
- Optimizing without profiling; "optimization" is just guessing
- Micro-optimizing non-hotspot code; spend time on the hottest functions
- Not considering cache effects; CPU cache profilers show real impact
- Benchmarking without warm-up; first runs hit different code paths
Further Reading
- Brendan Gregg, Systems Performance and Flame Graphs
- Java Flight Recorder documentation (Java)
- pprof documentation (Go)