Performance Optimization

Load this skill when profiling, optimizing, or reviewing code for performance — latency, throughput, memory, or CPU.

When to load: a change touches the request hot path (handler.rs, static_files/, compression.rs, fs/stream.rs), a benchmark is being added or interpreted, a regression in requests/sec or latency is suspected, or a new dependency may affect binary size or runtime cost.

General Approach

• Measure before optimizing: Use profiling tools (perf, flamegraph) to identify bottlenecks. Never optimize based on intuition
• Set a target: Define acceptable latency/throughput before starting. Stop optimizing when the target is met
• Optimize the hot path: Focus on code that runs on every request. Startup code and config parsing are low priority

Rust Performance

• Profile with perf and flamegraph: cargo flamegraph --bin static-web-server for CPU profiles. Profile under load (e.g., wrk or bombardier)
• Profile heap allocations with DHAT: Use DHAT or dhat-rs to find hot allocation sites. Reducing 10 allocations per million instructions can have measurable impact
• Avoid unnecessary allocations: Prefer &Path over &PathBuf, &[u8] over Vec<u8>, pass by reference where ownership is not needed
• Pre-compute at startup: Canonicalize paths, parse config, compile regex patterns, build Aho-Corasick automata once. Never on the request path
• Pre-allocate collections: Use Vec::with_capacity, String::with_capacity when the size is known
• Stream large responses: Use tokio::fs::File + tokio::io::copy for file serving. Never buffer the full file in memory
• Inline small hot functions: Use #[inline] on small functions called on every request (e.g., header name normalization, MIME lookups). Use #[cold] on error-path functions to guide branch prediction away from the hot path
• Prefer filter_map over filter().map(): Avoids an intermediate layer in hot iterator chains
• Use iter().copied() for small types: When iterating over &u8, &u32, etc., .copied() lets LLVM generate better code than receiving references
• Use chunks_exact when chunk size evenly divides length: Faster than chunks because it eliminates a remainder check per iteration
• Prefer ok_or_else over ok_or: ok_or(expensive()) always evaluates its argument. ok_or_else(|| expensive()) is lazy and only evaluates on None
• Eliminate bounds checks in hot loops: Use iteration instead of index-based access, or add an upfront assertion on the range to let the compiler prove bounds are safe

HTTP Performance

Connection Handling

• HTTP/1.1 keep-alive: Enabled by default via Hyper. Reduces connection setup overhead for subsequent requests
• HTTP/2 multiplexing: Enable with --http2 --tls. Multiple concurrent streams over a single TCP connection
• Worker threads: Default is num_cpus * 1. Increase --threads-multiplier for workloads with mixed CPU and I/O blocking (e.g., many concurrent clients with dynamic compression enabled — compression per-request is CPU-bound but high concurrency adds I/O wait interleaving). For pure CPU-bound workloads with minimal I/O, increasing threads beyond CPU count rarely helps.
• Max blocking threads: Default 512. For I/O-heavy patterns (large file serving), this is sufficient
• Graceful shutdown: Use --grace-period to allow in-flight requests to complete before shutdown

Compression Tradeoffs

• Static compression is free: Pre-compressed .br/.gz/.zst files are served with zero CPU. Always prefer this for production
• Dynamic compression overhead: On-the-fly compression trades CPU for bandwidth. Use --compression-level fastest for high-traffic sites
• Minimum size threshold: Responses below 860 bytes skip dynamic compression entirely — the overhead exceeds any bandwidth savings
• Compression algorithm priority (by compression ratio × speed): zstd > brotli > gzip > deflate. zstd offers the best ratio-speed tradeoff

Caching Headers

• Cache-Control is enabled by default: SWS sets max-age based on file extension:
  • 1 year for static assets (.css, .js, .png, .woff2, etc.)
  • 1 hour for feeds/API (.json, .xml, .rss, .atom)
  • 1 hour fallback for unknown extensions

You can override these defaults per file or extension using the configuration file. The above values are defaults, not hardcoded limits.

• Conditional requests: SWS supports If-Modified-Since and If-Unmodified-Since via ConditionalHeaders. Returns 304 when the file hasn't changed
• ETag not implemented: SWS uses Last-Modified instead. For byte-level cache validation, put SWS behind a CDN or reverse proxy

File I/O Performance

Buffering

• Optimal buffer size: optimal_buf_size() selects the best buffer size based on file metadata (uses std::fs::Metadata::blksize() when available)
• BufReader with take(): For byte-range requests, a BufReader wraps the file handle and limits bytes read to the requested range
• Streaming avoids full-file buffering: FileStream reads in chunks. The response body is a stream, not a byte buffer

Path Operations

• Canonicalize once at startup: The root directory is canonicalized in server/opts.rs. Per-request path resolution reuses this
• try_metadata() caches nothing: Each call (in src/fs/meta.rs) is a filesystem syscall. The experimental memory cache feature (mini-moka, in src/mem_cache/) caches file metadata and content
• Avoid clone() in the hot path: static_files.rs avoids cloning file paths for non-directory requests

Pre-compressed Static Files

• Zero-CPU serving: SWS detects .br/.gz/.zst variants via Accept-Encoding and serves them directly. No compression step runs
• Build-time pre-compression: Generate variants with maximum quality: brotli -q 11, gzip -9, zstd -19. SWS serves them as-is
• Vary header: Vary: Accept-Encoding is appended so caches know to store multiple variants

Memory

• Minimal per-connection state: SWS stores only the remote address and handler opts (shared via Arc). No per-connection buffers
• Response body is a stream: File contents are streamed, not buffered. Exception: small generated responses (health endpoint, error pages, directory listing HTML)
• Experimental in-memory cache: mini-moka (in src/mem_cache/) caches hot files in memory with LFU admission and LRU eviction. Configurable capacity (default 100 entries), ttl (default 1800s), tti, and max_file_size. Keys use CompactString to reduce allocation

Allocation Patterns

• Prefer clone_from over reassign-and-clone: a.clone_from(&b) reuses a's existing heap allocation when possible, avoiding an extra alloc/free. Especially valuable for Vec and String in hot loops
• Reuse collections across iterations: Declare the collection outside the loop, call .clear() at the end of each iteration. Avoids repeated alloc/free while keeping the heap allocation alive
• Use Cow<'_, str> / Cow<'_, Path> for mixed borrowed/owned data: Avoids allocating a String/PathBuf when the data is already a static literal or an existing slice that won't be modified
• Use SmallVec<[T; N]> for short, stack-like sequences: When most allocations hold ≤ N elements (e.g., header value lists, index file candidates), smallvec avoids heap allocation entirely for the common case
• Convert finalized Vec to Box<[T]> with into_boxed_slice(): Drops the unused capacity word, shrinking the type from 3 words to 2. Good for config-time data that is built once and never grown
• Return impl Iterator<Item=T> instead of Vec<T> from helpers: Avoids an allocation when the caller only needs to iterate
• Avoid format! when a literal or write! suffices: Every format! call allocates a String. Write directly to a &mut String or use std::fmt::Write instead

Type Sizes

• Keep hot types under 128 bytes: The compiler emits memcpy for values larger than 128 bytes. If a hot type exceeds this, check its layout with RUSTFLAGS=-Zprint-type-sizes cargo +nightly build --release
• Box large enum variants: If one variant is much larger than the others, box its fields to bring all variants to a similar small size. Reduces stack pressure and cache churn
• Use smaller integer types for index/count fields: Prefer u32 over usize for counts and offsets stored in frequently instantiated structs (e.g., header tables, path segments). Cast to usize at use sites
• Assert type sizes in tests: Add static_assertions::assert_eq_size!(HotType, [u8; N]); for performance-critical types so that accidental size regressions cause a compile error

Release Build Configuration

The default cargo build --release profile is a good starting point, but the following options can improve throughput for production SWS builds:

Option	Effect	Cargo.toml
lto = "thin"	Cross-crate inlining, 5–15% speedup, moderate compile cost	[profile.release]
codegen-units = 1	Single codegen unit, enables more optimizations, slower compile	[profile.release]
panic = "abort"	Removes unwinding machinery, smaller binary, slight speedup	[profile.release]

For a custom server build where broad CPU compatibility is not required:

RUSTFLAGS="-C target-cpu=native" cargo build --release

This emits AVX/SSE instructions optimal for the build machine, which can improve compression throughput.

Note: target-cpu=native produces a non-portable binary. Do not use for distributed release artifacts.

Benchmarking

Micro-benchmarks (CodSpeed)

The benches/ directory is a standalone crate with Criterion benchmarks for hot-path functions (path sanitization, header handling, redirects, basic auth). They run on every pull request via the codspeed workflow and are tracked on CodSpeed.

cd benches

# Plain Criterion run (local timings only)
cargo bench

# Same suite measured with CodSpeed CPU simulation (deterministic, flamegraphs)
cargo codspeed build
codspeed run --mode simulation -- cargo codspeed run

Add a bench when a change touches the request hot path so a future regression is caught by CI instead of by users.

Tools

• HTTP load generators: wrk, bombardier, oha, hey
• CPU profiling: perf record + flamegraph, cargo flamegraph, cargo instruments (macOS), samply (cross-platform)
• Allocation profiling: dhat-rs (all platforms), DHAT via Valgrind (Linux) — identifies hot allocation sites
• Monitoring: Prometheus metrics via --metrics + /metrics endpoint

What to Measure

• Requests per second at concurrency levels: 1, 10, 100, 1000
• Latency percentiles: p50, p95, p99
• Memory usage: RSS before and under load
• CPU utilization: Per-core usage during load test
• Allocation rate: Use DHAT to confirm per-request allocations are not growing unexpectedly

Checklist

• Is there a benchmark or profile showing the bottleneck?
• Are paths canonicalized once, not per-request?
• Is static compression used where possible (zero CPU)?
• Is dynamic compression size-threshold applied (860 bytes)?
• Are large files streamed, not buffered?
• Are Cache-Control headers set appropriately for the content type?
• Are worker threads configured for the workload?
• Is keep-alive or HTTP/2 enabled for connection reuse?
• Are hot allocation sites identified with DHAT or similar?
• Are collections pre-allocated or reused rather than recreated per request?
• Are clone() calls on the hot path justified — or replaceable with clone_from, Cow, or a reference?
• Are hot types under 128 bytes (no unintended memcpy)?
• Are #[inline] / #[cold] attributes applied where profiling shows they help?
• Are ok_or_else / lazy combinators used instead of eager alternatives in hot paths?