Jolteon: Unleashing the Promise of Serverless for Serverless Workflows
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
An orchestrator called Jolteon implements a stochastic performance model combining analytical whitebox formulas and blackbox variability, formulates a chance-constrained optimization (confidence level 99.9%), and uses sample approximation plus convexity to solve for per-stage function size and parallelism. A Python prototype deploys on AWS Lambda, profiles runs, then computes configurations via gradient descent and a probe step. Evaluated on ML pipeline, video analytics and TPCs query workflows, it yields up to 2.3× cost savings and 2.1× lower latency versus Orion.
AI-written summary · about
chapters (12)
- 0:00Jolteon and serverless workflows
- 1:24Serverless workflow challenges
- 3:12Limitations of performance models
- 6:23Jolteon’s solution approach
- 7:06Stochastic performance model
- 11:22Optimization problem formulation
- 13:50Stochastic optimization algorithm
- 14:34Prototype and evaluation setup
- 15:17Comparison with baselines
- 16:20Latency and cost results
- 16:41Performance gains summary
- 17:02Conclusion and future
related talks
21mAI Training Orchestration Evolution with Serverless Building Blocks
28mServerless Computing for Mathematical Optimization
35mRADON: DevOps framework to optimally exploit serverless computing tech...
28mThe Future of Serverless
40mThe State of Serverless Computing
42mNo Server Is Better Than No Server: Building Blocks for Serverless Architectures
46mServerless for Data Scientists
29mGoing Serverless