Shockwave: Fair and Efficient Cluster Scheduling for Dynamic Adaptation in Machine Learning
summary
Shockwave is a cluster scheduler that predicts batch‑size and throughput changes using a Bayesian mixture‑of‑regimes model and expands Fisher markets into a volatile Fisher market with a Gurobi solver to plan allocations across future rounds. The implementation handles gradient‑noise and accordion adaptive patterns, runs trace and cluster experiments on Gavel and Microsoft Philly traces, and reports a 1.3× reduction in makespan and 2× improvement in finish‑time fairness versus prior fair schedulers.
AI-written summary · about
chapters (7)
- 0:00Introduction to cluster scheduling
- 1:30Dynamic workload adaptation
- 3:41Challenges in scheduling
- 6:13Shockwave system overview
- 10:23Bayesian predictive model
- 13:26Market-based allocation
- 15:52Results and conclusion
related talks
12mEasy and affordable access to GPUs for AI/ML workloads
05mProject Lightning Talk: Scheduling AI Workload Among Multiple Clusters
24mVolcano – Cloud Native Batch System for AI, BigData and HPC
16mSLearn - A Case for Task Sampling based Learning for Cluster Job Scheduling
26mOptimal Approaches for Real-Time Machine Learning with Apache Spark on Kubernetes.
10mFair Scheduling for Deep Learning Workloads in Kubernetes
01h 03mScaling up machine learning without tears (and what do programming languages ha...
22mThemis: Fair and Efficient GPU Cluster Scheduling