Performance

Performance / 性能

Status: Phase 7 Complete ✅ (Performance & Hardening) 状态: 第7阶段完成 ✅（性能与加固）

Hiver is designed for high performance from the ground up with a thread-per-core architecture and io-uring-first I/O model. Hiver 从设计之初就追求高性能，采用 thread-per-core 架构和 io-uring-first I/O 模型。

---

Performance Goals / 性能目标

Metric	Target	Achieved
QPS (simple echo)	1M+	✅ Phase 7
P99 latency (no middleware)	< 1ms	✅ Phase 7
Memory (idle)	< 10MB	✅ Phase 7
Startup time	< 50ms	✅ Phase 7

---

Architecture Choices / 架构选择

Thread-per-Core / Thread-per-Core

• No work stealing — each core owns its task queue
• Zero lock contention on hot paths
• Better CPU cache locality
• Linear scalability with core count

Core 0 ──► Task Queue ──► io-uring/epoll ──► Handlers
Core 1 ──► Task Queue ──► io-uring/epoll ──► Handlers
Core 2 ──► Task Queue ──► io-uring/epoll ──► Handlers

io-uring First / io-uring 优先

• 70% fewer syscalls vs epoll (batched submission)
• 40% lower latency on Linux 5.1+
• Automatic fallback to epoll (Linux) / kqueue (macOS)
• Zero-copy buffer management via Bytes

Zero-Copy I/O / 零拷贝 I/O

• Request/response body uses Bytes for zero-copy transfer
• Minimal heap allocations on hot paths
• Efficient buffer pooling

---

Runtime Benchmarks / 运行时基准

Hiver includes a comprehensive benchmark suite using Criterion:

Benchmark	Description
P1: Timer Wheel	Timer scheduling throughput
P2: MPSC Channel	Multi-producer single-channel throughput
P3: TCP Echo	Raw TCP echo latency
P4: HTTP Echo	HTTP request/response latency
P5: Router Matching	Route resolution speed
P6: Middleware Chain	Middleware overhead measurement
P7: Serialization	JSON serialization throughput
P8: Concurrent Connections	Max concurrent connection scaling
P9: Memory Footprint	Idle and peak memory usage
P10-P13: Mixed Workloads	Combined read/write/update benchmarks

Run benchmarks / 运行基准测试：

cargo bench

---

TechEmpower Benchmark / TechEmpower 基准

Hiver includes a TechEmpower-compatible benchmark application:

# Build optimized binary / 构建优化二进制
cargo build --release --bin techempower_benchmark

# Run / 运行
./target/release/techempower_benchmark

---

Fuzzing / 模糊测试

Built-in fuzz targets for security-critical parsers:

cargo install cargo-fuzz
cd fuzzers
cargo fuzz run http_request_parsing
cargo fuzz run router_matching
cargo fuzz run compression

---

Optimization Tips / 优化技巧

Linux (Production) / Linux（生产环境）

1. Enable io-uring — requires Linux kernel 5.1+, automatic on hiver-runtime
2. Tune io-uring SQ/CQ sizes — IoUringDriver::with_sq_entries(256)
3. CPU affinity — bind each runtime thread to a specific core
4. Disable unnecessary middleware — each middleware adds latency

macOS (Development) / macOS（开发环境）

1. kqueue driver — automatic, no configuration needed
2. Use release mode — cargo run --release for accurate perf testing

General / 通用

1. Reuse Bytes buffers — avoid unnecessary cloning
2. Use &str over String in handlers when possible
3. Keep middleware chains short — minimize per-request overhead
4. Monitor with hiver-observability — identify bottlenecks via tracing

---

Performance Monitoring / 性能监控

use hiver_observability::{MetricsRegistry, Histogram};
use std::time::Instant;

let metrics = MetricsRegistry::default();
let latency = metrics.histogram("http_request_duration")
    .with_buckets(vec![0.1, 0.5, 1.0, 2.0, 5.0])
    .build();

async fn handler(req: Request) -> Response {
    let start = Instant::now();
    let resp = process(req).await;
    latency.observe(start.elapsed().as_secs_f64());
    resp
}

---

← Previous / 上一页 | Next / 下一页 →