Harmonizing Efficiency and Practicability: Optimizing Resource Utilization in...
summary
Jiagu targets serverless resource waste with two concrete techniques: pre-decision scheduling builds per-node capacity tables that answer most placement queries instantly (fast pass) using a 56% replication observation, falling back to slow prediction only for rare functions; dual-staged scaling adds a releasing phase that turns idle instances into cached instances and uses 1 ms logical code starts to avoid cold starts. Prototype on real cloud traces shows 54.8% higher deployment density vs Kubernetes, 81.0–93.7% lower scheduling cost, and 57.4–69.3% less cold-start latency while maintaining QoS.
AI-written summary · about
chapters (9)
- 0:00Serverless resource optimization
- 3:16Low resource utilization causes
- 7:32Challenges in prior approaches
- 10:44JAGU’s design
- 16:53Capacity updates and scheduling
- 17:14Stage scaling for autoscaling
- 19:25Evaluation results and gains
- 20:30Conclusion and future work
- 20:51Q&A
related talks
10mEnergy-Efficient and Cost-Optimized Serverless Computing
27mServerless Performance on a Budget
18mHelp Rather Than Recycle: Alleviating Cold Startup in Serverless Computing...
33mGoing FaaSter: Cost-Performance Optimizations of Serverless on Kubernetes
21mWhat is the Future of Serverless Computing?
27mServerless on Kubernetes: WASM, Knative, or Regular Autoscaling – What Actually Works?
01h 02mServerless Architecture Patterns
43mBest Practices for Serverless Developers