ALPS: An Adaptive Learning, Priority OS Scheduler for Serverless Functions
summary
ALPS unifies approximate SRPT via user-space SRPT simulation with eBPF hooks to CFS in a decoupled frontend/backend: the user-space front end learns time-slice upper bounds and per-function priorities from recent traces, writes policies to eBPF maps, and the kernel-side eBPF alters CFS to grant extra quanta or override RB-tree ordering. Prototype modifies OpenLambda and Docker with a few lines, evaluates on Huawei and Azure FaaS traces, and reports a 57.2% reduction in average function execution duration.
AI-written summary · about
chapters (6)
- 0:00ALPS scheduler intro
- 2:10CFS scheduler limits
- 6:28Adaptive learning priority design
- 12:13ALPS implementation
- 15:46Performance results
- 17:57Conclusion
related talks
22mUnFaaSener: Latency and Cost Aware Offloading of Functions from Serverless Platforms
19mSponge: Fast Reactive Scaling for Stream Processing with Serverless Frameworks
36mFunctional Programming in Serverless World
17mTowards High-Performance Transactional Stateful Serverless Workflows with...
51mFunction as a Service in private cloud
38mBackground Processing, Serverless Style
21mAI Training Orchestration Evolution with Serverless Building Blocks
16mFork in the Road: Reflections and Optimizations for Cold Start Latency in Production...