Cook: A Fair Preemptive Resource Scheduler for Compute Clusters

This post introduces Cook, a fair preemptive resource scheduler for compute clusters. It details the problem of balancing interactive and batch job requirements, proposes a Cumulative Resource Share (CRS) metric for fairness and prioritization, and explains how preemption is used to reclaim resources for higher-priority tasks. Benchmarks using a production trace demonstrate that Cook effectively reduces latency for interactive jobs while maintaining high cluster utilization.