Can't Be Late: Optimizing Spot Instance Savings under Deadlines
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
Proposes a parameter-free policy Uniform Progress that switches to on-demand only when observed progress falls below the time-uniform expectation, and implements it in the SkyPilot broker SkyPilot to handle preemptions and changeover delays. Evaluated on 3-month AWS V100 traces and ILP-derived omniscient bounds, it halves the greedy policy’s gap to optimal, typically stays within 10% of optimal, and yields real-workload savings of 27%-84% across deadlines and cluster sizes.
AI-written summary · about
chapters (7)
- 0:00Introduction to spot instances
- 1:25The challenge of availability
- 4:22Policy formulation
- 7:39Developing a robust strategy
- 10:08Advanced time-slice policy
- 13:28Evaluation and results
- 16:18Future directions and conclusion
related talks
32mBlending Spot and On-Demand Instances and Optimizing Costs
19mStarburst: A Cost-aware Scheduler for Hybrid Cloud
05mCutting Cloud Costs in COVID-19!
17mSkyPilot: An Intercloud Broker for Sky Computing
16mOptimizing Network Provisioning through Cooperation
42mAWS Spot instances: huge savings (or huge risks?)
11mUnlocking the Power of Machine Learning for Cloud Workload Opt.
43mSpot Instances