Stochastic Page Placement

Abstract

Imperfect virtual-to-physical page mappings cause excess conflict misses in direct-mapped CPU caches. Perfect place ment minimizes the conflict miss rate and is useful for determining an upper bound on direct-mapped cache performance. In this paper, we introduce a polynomial-time alternative to perfect placement called stochastic page placement. Our method combines a general optimization technique with trace-driven simulation to find a local minimum of the perfect page placement solu tion space. Two stochastic placement policies are presented: GA placement uses a genetic algorithm to compute virtual page mappings, and SA placement uses simulated annealing. Parallel distributed simulation algorithms for perfect placement and GA placement are also presented. For our workloads, GA placement generally outperforms SA placement, but other careful placement policies that exploit workload properties usually outperform stochastic placement.