Performance Optimization

Core Question

Have you measured first, or are you guessing?

Intuition about performance is often wrong. The code you think is slow frequently isn't. Profile first, then optimize the actual bottleneck with data-driven decisions.

Quick Decisions

Optimization Priority

Optimize in this order — each level has roughly 10x less impact than the one above:

1. Algorithm & data structure — O(n) vs O(n^2) dwarfs everything else
2. Data layout — SoA vs AoS, cache-friendly access, avoid pointer chasing
3. Allocations — with_capacity, reuse buffers, arena allocators
4. Compiler hints — LTO, PGO, codegen-units, inline hints
5. SIMD & low-level — portable SIMD, bounds check elimination, target-cpu

Profile BEFORE Optimizing

Tools

Workflow

1. Write correct code first
2. Write benchmarks for suspected hot paths
3. Profile under realistic load
4. Identify actual bottlenecks (top 10% of time)
5. Optimize ONE thing
6. Measure improvement with same benchmark
7. Repeat if needed

Criterion Quick Setup

use criterion::{black_box, criterion_group, criterion_main, Criterion};

fn bench_hot_function(c: &mut Criterion) {
    let data = generate_test_data(1000);
    c.bench_function("hot_function", |b| {
        b.iter(|| hot_function(black_box(&data)))
    });
}

criterion_group!(benches, bench_hot_function);
criterion_main!(benches);

Anti-Patterns to Watch For

Usage Scenarios

Scenario 1: "My web handler is slow and I'm not sure why" -> Run cargo flamegraph on a representative workload. Look for wide bars (time hogs), malloc/free (allocation heavy), memcpy (unnecessary copies). Optimize only what the flamegraph shows as hot.

Scenario 2: "I have a Vec that I fill in a loop and it's showing up in profiling" -> Check if size is known: use with_capacity(). If the Vec is reused across iterations: .clear() instead of creating new. If always small (< 8 items): consider SmallVec. If fixed after creation: convert to Box<[T]>.

Scenario 3: "I need maximum throughput for a release binary" -> Set release profile: lto = "fat", codegen-units = 1, panic = "abort", strip = true. For deployment on known hardware: RUSTFLAGS="-C target-cpu=native". For maximum gains: use PGO with representative workloads.

Release Profile

[profile.release]
opt-level = 3
lto = "fat"
codegen-units = 1
panic = "abort"
strip = true

[profile.bench]
inherits = "release"
debug = true
strip = false

[profile.dev.package."*"]
opt-level = 3

Reference Files

Cross-References