SONIC: Application-aware Data Passing for Chained Serverless Applications
summary
Compares three concrete data-passing modes VM-Storage, Direct-Passing, Remote-Storage and shows no single mode wins: small fan-out favors VM-Storage, medium fan-out favors Direct-Passing, high fan-out or low VM bandwidth favors Remote-Storage. Builds SONIC, a manager integrated with Open-Lambda on Amazon EC2, that profiles DAGs, uses simple regression models and a Viterbi optimizer to pick per-edge method and placement. Evaluated on three analytics apps (video, LightGBM, MapReduce), 50–60 concurrent jobs, 60 YouTube videos, yields 1.8× speedup vs SAND and approaches an oracle within 19%.
AI-written summary · about
chapters (15)
- 0:00Intro
- 0:27Background: Serverless Computing DAG & Data Passing
- 1:52Serverless Data-Passing Challenges
- 2:32SONIC's Data-Passing Alternatives
- 4:13Data-passing performance trade-off
- 5:02Direct-Passing vs Remote Storage With higher network bandwidth, the crossover point between Direct passing and Remote-passing shifts to higher fanout values
- 5:58OUR SOLUTION: SONIC Hybrid Data-Passing Approach
- 7:39Greedy Data-Passing Decisions: Pitfalls
- 9:05SONIC: Design Overview
- 10:06Evaluation: Baselines
- 11:13Evaluation: E2E Latency and Cost
- 13:00Content Sensitivity (Cont.)
- 13:41Evaluation: Content Sensitivity
- 14:29Evaluation: Scalability
- 15:17Conclusion
related talks
21mHarmonizing Efficiency and Practicability: Optimizing Resource Utilization in...
40mThe State of Serverless Computing
49mBuilding Next-Generation Applications with EDA
26mShuffling, Fast and Slow: Scalable Analytics on Serverless Infrastructure
20mDataflow Engines for Executing Cloud Applications: a Maslow's Hammer or Natural Outcome?
31mAccelerate business innovation with serverless applications
41mTechnology mix: Around the serverless in 180 days
19mSponge: Fast Reactive Scaling for Stream Processing with Serverless Frameworks