Improving Transfer Times in the Backbone Network Using QUIC Jump Start
summary
Meta deploys QUIC jumpstart with 0-RTT and a modified congestion controller that keeps a per-process EWMA of cwnd and min RTT per region, sets a pacing rate, and reuses that state to seed new connections. Results: requests <1 MB run ~10% faster than Meta’s double-proxy pooling; initial tests showed ~30% tail-slower at ~8 MB due to QUIC CPU bottlenecks, but tuning ACK frequency and enabling GRO matches proxy performance. Ongoing issues: path-awareness, dampened loss signals, and possible centralization.
AI-written summary · about
chapters (12)
- 0:00QUIC Jump Start intro
- 1:08TCP/TLS handshake overhead
- 3:15Data transfer phases
- 5:43Slow start vs jump start
- 6:48Proxy servers and pooling
- 8:18QUIC Jump Start design
- 10:21QUIC Zero RTT results
- 11:43CPU bottlenecks in QUIC
- 13:29Bandwidth and latency impact
- 13:52QUIC Jump Start challenges
- 16:20Simplifying backbone with QUIC
- 17:22Summary and plans
related talks
19mLive Media Over QUIC
38mWebRTC and QUIC: how hard can it be?
17mQDSR: Accelerating Layer-7 Load Balancing by Direct Server Return with QUIC
54mQUIC illustrated
42mMQTT over QUIC: Opportunities and Challenges
22mQ&A Session 2 – Bharat Parekh
35mOn QUIC, HTTP/3 and tech stack unification
31mRapid Start: Faster Internet Connections, with Ruby's Help