Overview
Explore a conference talk that delves into Shockwave, an innovative cluster scheduling system designed for dynamic adaptation in machine learning. Learn how this scheduler addresses the challenges of fairness and efficiency in ML training environments where model structure and hyperparameters are dynamically adjusted. Discover the key ideas behind Shockwave, including its extension of market theory to dynamic settings and the use of stochastic dynamic programming. Understand how Shockwave outperforms existing fair scheduling schemes, improving makespan and fairness for ML jobs with dynamic adaptation. Gain insights into the future of cluster scheduling for machine learning workloads and its potential impact on accelerating training without sacrificing accuracy.
Syllabus
NSDI '23-Shockwave: Fair and Efficient Cluster Scheduling for Dynamic Adaptation in Machine Learning
Taught by
USENIX