Rust Performance

Rust Performance Patterns

Profiling Tools

# Flamegraph (CPU profiling)
cargo install flamegraph
cargo flamegraph --bin my-server

# Benchmarks
cargo install cargo-criterion
cargo criterion

# Inspect generated assembly
cargo install cargo-asm
cargo asm my_crate::hot_function

# Memory profiling with dhat
# Add dhat feature to Cargo.toml and instrument main

Iterator Performance

Iterators compile to the same machine code as hand-written loops — zero overhead.

// Single-pass chain — no intermediate allocations
let result: Vec<_> = data.iter()
    .filter(|&&x| x > 0)
    .map(|&x| x * 2)
    .collect();

// Avoid collect() in the middle of a chain
// BAD: two allocations
let filtered: Vec<_> = data.iter().filter(|&&x| x > 0).collect();
let mapped: Vec<_> = filtered.iter().map(|&&x| x * 2).collect();

Avoiding Allocations

// &str instead of String when ownership isn't needed
fn process(input: &str) -> usize { input.len() }

// Stack allocation for small collections
use smallvec::SmallVec;
let mut v: SmallVec<[u8; 16]> = SmallVec::new(); // on stack until > 16 bytes

// Cow for conditionally owned data
use std::borrow::Cow;
fn normalize(s: &str) -> Cow<str> {
    if s.chars().all(|c| c.is_lowercase()) {
        Cow::Borrowed(s)
    } else {
        Cow::Owned(s.to_lowercase())
    }
}

// Pre-allocate known sizes
let mut result = Vec::with_capacity(items.len());
for item in &items { result.push(transform(item)); }

Memory Layout Optimization

// Largest fields first to minimize padding
#[derive(Debug)]
struct Optimized {
    b: u64,  // 8 bytes
    d: u32,  // 4 bytes
    a: u8,   // 1 byte
    c: u8,   // 1 byte
    // 2 bytes padding
}  // 16 bytes total

// Compare to field-order that wastes space
#[derive(Debug)]
struct Wasteful {
    a: u8,   // 1 byte + 7 padding
    b: u64,  // 8 bytes
    c: u8,   // 1 byte + 3 padding
    d: u32,  // 4 bytes
}  // 24 bytes total

// Explicit layout control
#[repr(C)]          // C-compatible layout
#[repr(align(64))]  // cache-line aligned
struct CacheAligned { data: [u8; 64] }

Cache-Friendly Data Structures

// Structure of Arrays (SoA) — better for bulk operations than Array of Structs
struct Particles {
    x: Vec<f32>,
    y: Vec<f32>,
    z: Vec<f32>,
    mass: Vec<f32>,
}

fn update_x(p: &mut Particles, velocities: &[f32], dt: f32) {
    // Sequential memory access = cache-friendly
    for (x, &vx) in p.x.iter_mut().zip(velocities.iter()) {
        *x += vx * dt;
    }
}

Parallelism with rayon

use rayon::prelude::*;

let sum: i64 = data.par_iter().map(|&x| expensive(x)).sum();

let mut v: Vec<i32> = (0..1_000_000).collect();
v.par_sort();

String Performance

// write! to a pre-allocated String instead of format! in hot paths
use std::fmt::Write;
let mut s = String::with_capacity(64);
write!(s, "user_{}_event_{}", user_id, event_id).unwrap();

// smol_str: stack-allocated for short strings
use smol_str::SmolStr;
let s: SmolStr = "short string".into(); // no heap for ≤23 bytes

Release Profile Tuning

[profile.release]
opt-level = 3
lto = "thin"       # link-time optimization
codegen-units = 1  # cross-function optimization
strip = true
panic = "abort"    # no unwinding overhead

[profile.bench]
inherits = "release"
debug = true       # keep symbols for profiling

Common Anti-Patterns

• clone() in hot loops — profile first; often avoidable with references or restructuring
• String where &str suffices — unnecessary heap allocation
• dyn Trait in hot paths — use generics (monomorphization) for zero-cost dispatch
• Blocking I/O on async runtimes — use tokio::task::spawn_blocking
• Profiling debug builds — always profile --release builds; debug is 10x slower