Fork in the Road: Reflections and Optimizations for Cold Start Latency in Production...
summary
Platform traces show 19.5% of requests are cold starts and 18% of invocations finish within 221 ms, exposing three overlooked bottlenecks: control path latency, resource contention, and user code initialization latency. AFaaS implements a short function runtime interface, fork-friendly preallocation/pooling for namespaces and cgroups, and a multi-level code-set fork tree. Deployed 18 months, it drops cold-starts to single-digit milliseconds (stable ~5.45–9.4 ms), reports 1.8×–8.1× end-to-end speedups, and eliminates hour-long degradation under high concurrency.
AI-written summary · about
chapters (9)
- 0:00Cold start latency intro
- 2:27Serverless container performance gaps
- 4:18Cold start three phases
- 6:07Control path latency inefficiencies
- 7:35Resource contention tail latency
- 9:44User code initialization fork design
- 11:11AAS performance evaluation
- 13:01Memory usage production results
- 14:26Conclusion future directions
related talks
33mGoing FaaSter: Cost-Performance Optimizations of Serverless on Kubernetes
27mServerless Performance on a Budget
23mCF Server-Less: CloudFoundry as a Platform for Server-less Computing
37mGo FaaSter: Cold Start Optimization in a Serverless Platform
37mGo FaaSter: Cold Start Optimization in a Serverless Platform
54mHow to avoid common mistakes and misconceptions when working...
25mCF Serverless: CloudFoundry as a Platform for Serverless Computing
53mHigh performance Serverless Java on AWS