Performance Profiling and Benchmarking

Performance work in Elixir should follow a repeatable loop:

1. measure,
2. identify bottleneck,
3. change one variable,
4. measure again.

Without this loop, optimization often creates complexity without meaningful gains.

Benchmarking with Benchee

Use Benchee for controlled micro/mid-level comparisons:

Benchee.run(%{
  "enum_map" => fn -> Enum.map(1..100_000, &(&1 * 2)) end,
  "for_comp" => fn -> for x <- 1..100_000, do: x * 2 end
})

Benchmark tips:

• warm-up sufficiently,
• test realistic input sizes,
• isolate noisy external dependencies.

Profiling for Hotspots

Profilers answer where time is spent, not just which function is faster in isolation.

Useful tools:

• :fprof for call-time analysis,
• :eprof for function-level profiling,
• tracing and telemetry for production paths.

BEAM-Specific Performance Signals

• scheduler utilization imbalance,
• mailbox growth and message queue pressure,
• excessive process churn,
• binary memory retention.

Inspect these before rewriting algorithms prematurely.

Optimization Priorities

• reduce unnecessary allocations/copies,
• batch expensive IO,
• move repeated computation out of hot loops,
• choose appropriate concurrency boundaries.

# cProfile + pytest-benchmark + tracing stacks
# Similar measurement-first methodology.

// Node perf hooks + clinic/flame tools
# Similar hotspot and throughput analysis approach.

# Benchee + BEAM profilers + Telemetry
# Combined offline and runtime profiling for reliable optimization.

Exercise

FAQ and Troubleshooting

Why do microbenchmarks improve but endpoint latency does not?

The real bottleneck may be IO, contention, or downstream services not represented in the microbenchmark.

How do I avoid over-optimizing?

Set target SLOs first. Stop optimizing when targets are met and complexity costs outweigh gains.

Should I optimize memory or CPU first?

Optimize the limiting resource observed in production metrics for your actual workload.