QUIC Transport Optimizations from a Reimplementation

Published on 2025-12-05

Reimplementing QUIC in C++ gave me room to tune congestion control and pacing. Here’s what moved the needle and why packet captures mattered.

What I Built

• C++ QUIC transport with custom congestion control and pacing
• gRPC harness for cross-language benchmarks
• Wireshark tracing inside a Docker-based latency simulator

Results

• ~25% reduction in average RTT under mixed loads
• Smoother cwnd evolution and fewer bursty loss events

Key Lessons

• Fine-grained pacing tames burst loss more than conservative cwnd alone.
• Benchmark harnesses need predictable network conditions; Docker tc profiles helped.
• Wireshark filters (udp.port && quic) made iterative tuning faster.

Next Steps

• Add BBR-style probing
• Bake in automated pcap diffing in CI