Starburst: A Cost-aware Scheduler for Hybrid Cloud
summary
Starburst is a meta-scheduler that assigns per-job waiting budgets, uses compute-weight (or its practical approximation star-weight with preemption) plus out-of-order scheduling to reserve capacity for large long jobs while packing small short jobs, driving up on-prem cluster utilization. Implemented with SkyPilot and Kubernetes autoscaling, evaluated on a 32‑GPU lab cluster and 512‑GPU simulations with 100k-job Philly and Helios traces, it reports 54–91% cloud cost reduction while increasing average JCT by at most ~5.8%.
AI-written summary · about
chapters (11)
- 0:00Starburst scheduler intro
- 1:06Sky computing overview
- 2:09Cost vs scheduling trade-off
- 6:25Starburst design approach
- 8:55Compute weight policies
- 11:25General Starburst method
- 13:0832-GPU cluster evaluation
- 14:10Simulations and trace comparisons
- 16:58Pareto curve optimization
- 17:18Data gravity impact
- 18:00Summary
related talks
05mSponsored Keynote: The Cloud (And Your $$$) is Not Infinite: Dynamic Scheduling at Every Layer
10mLightning Talk: Does Cloud Elasticity Pay off for HPC Workloads?
33mUncover Kubernetes cost-optimization trends
17mSkyPilot: An Intercloud Broker for Sky Computing
47mCloud Native and Sustainable, Reproducible Scientific Computing
31mAccelerate AI adoption: automated cloud efficiency
33mKubernetes cost optimization - How to actually do it
16mSky Computing: Charting a New Course in AI Cloud Infrastructure