Prism: Proxies without the Pain
summary
Prism moves established connections from front-end to backend using TCP handoff and a two-phase handoff that blocks client packets while the switch rules update, implemented with eBPF and Linux connection serialization. Measured handoff cost is about 232 μs handoff latency, benchmarks with 100 parallel TCP connections and nginx frontends show backend CPU and network become fully utilized, yielding up to 3.7× throughput (3.4× with TLS) compared to traditional frontend proxy architectures.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:28Oceanstage
- 0:51Frontend
- 4:12Metrics
- 5:00Results
- 5:47Improvements
- 6:39Contribution
- 7:59TCP Handler
- 10:16End of Cost
- 10:54Conclusion
related talks
34mHow Many Proxies Do You Need?
31mUnleashing the Tides of Kubernetes Networking by Removing kube-proxy
45mHow the Network Protocols You Choose Ultimately Affect Your Applications
24mHow We Built Protosockets to Go Beyond HTTP and gRPC
18mHigh Performance Service Meshes with gRPC Proxyless
23mSYN Proxy at Scale with BPF
20mPostMan: Rapidly Mitigating Bursty Traffic by Offloading Packet Processing
42mAccelerating Envoy with the Linux Kernel