Serverless Performance on a Budget
summary
Serverless platforms still see noticeable cold starts—measured around 200 ms cold starts on major providers—and performance gains depend on trade-offs, not magic. The presentation shows concrete optimizations: function reuse (shared instances), a function runtime pool to cut deployment time in half, hierarchical prefetching of large function artifacts, and predictive pre-warming using workflow graphs (example: Fission Workflows pre-warms downstream functions). Each gains latency at the cost of extra idle CPU, memory or storage.
AI-written summary · about
chapters (22)
- 0:00Introduction
- 0:18Performance Model
- 1:16Cost Model
- 3:56Runtime Components
- 6:34Performance
- 7:32Cold Start
- 8:26Techniques
- 8:48First Optimization
- 10:56Tradeoff
- 11:34Sharing
- 11:47Insight
- 12:33Research pooling
- 13:20Rebalancing
- 13:39Research Pool
- 14:42Prefetching
- 15:53Hierarchy of Layers
- 17:30Function Prewarming
- 18:00Visualizing Function Prewarming
- 19:29Static Information
- 21:27fission workflows
- 22:49prewarm
- 24:26Summary
related talks
37mGo FaaSter: Cold Start Optimization in a Serverless Platform
37mGo FaaSter: Cold Start Optimization in a Serverless Platform
33mGoing FaaSter: Cost-Performance Optimizations of Serverless on Kubernetes
35mEfficiency of Serverless Architectures with AWS Solutions
30mPerformance Tales of Serverless
23mServerless in the Wild: Characterizing and Optimizing the Serverless Workload
57mTackling Dreaded .Net Cold Starts When Going Serverless
43mBest Practices for Serverless Developers