Skip to content

BENCHMARK.md

Latest commit

 

History

History
99 lines (72 loc) · 2.7 KB

BENCHMARK.md

File metadata and controls

99 lines (72 loc) · 2.7 KB
Raw

Benchmarks

HTTP/3 request/response throughput on localhost (macOS M-series, single-threaded servers).

Quick Start

# Build release
zig build -Doptimize=ReleaseFast

# Start server
zig-out/bin/server --port 4434

# Run bench (another terminal)
zig-out/bin/bench --port 4434 -c 50 -n 100

Bench Options

zig-out/bin/bench [options]
  --port PORT        Server port (default: 4434)
  -c, --connections  Number of sequential connections (default: 1)
  -n, --requests     Requests per connection (default: 100)
  -z, --zerortt      Enable 0-RTT session resumption

Server Options

zig-out/bin/server [options]
  --port PORT        Listen port (default: 4434)

Results (macOS M-series, localhost, 50 conn × 100 req)

Server Comparison (Zig bench client)

Server Req/s Handshake p50 p99
quic-go (Go) 13,278 876µs 264µs 733µs
quic-zig (Zig) 11,009 1,030µs 404µs 586µs
quiche+mio (Rust) 1,114 1,143µs 329µs 1,992µs

Client Comparison (quic-zig server)

Client Req/s p50 p99
Zig bench 11,009 404µs 586µs
Go bench 6,085 136µs 1,722µs

0-RTT Handshake Speedup

Mode Handshake Req/s
1-RTT 932µs 11,258
0-RTT 410µs 12,409

Debug vs Release

Build Req/s Handshake p50
Debug 3,561 5,441µs 1,411µs
Release 11,678 891µs 383µs

Reproducing Comparisons

quic-go server

cd interop/quic-go
go build -o h3server_bin ./h3server
./h3server_bin --addr localhost:4435 \
  --cert ../certs/server.crt --key ../certs/server.key

quiche server (Rust + mio)

cd interop/quiche
cargo build --release --bin h3server
./target/release/h3server --addr 127.0.0.1:4436 \
  --cert ../certs/server.crt --key ../certs/server.key

Go bench client

cd interop/quic-go
go build -o bench_bin ./bench
./bench_bin -port 4434 -c 50 -n 100

Notes

  • All servers use the same TLS certificates (interop/certs/)
  • quiche's standalone server processes connections serially; in Cloudflare's production stack it uses connection-per-worker sharding
  • The Go client shows lower p50 latency but lower throughput due to net/http overhead
  • For multi-core scaling, run multiple server instances behind a QUIC-aware load balancer (see QUIC-LB)