Cache-Related Preemption and Migration Delays: Empirical Approximation and Impact on Schedulability

A job that is impeded by a preemption or migration incurs additional cache misses when it resumes execution due to a loss of cache affinity. While often regarded negligible in scheduling-theoretic work, such cache-related delays must be accounted for when comparing scheduling algorithms in real systems. Two empirical methods to approximate cache-related preemption and migration delays on actual hardware are proposed, and a case study reporting measured average- and worst-case overheads on a 24-core Intel system with a hierarchy of shared caches is presented. The widespread belief that migrations are always more costly than preemptions is refuted by the observed results.