An adaptive page replacement scheme for scientific applications

Abstract

© 2020 IEEE.Memory management is playing an increasingly important role in application performance owing to the rapid increase in the amount of data utilized in modern computing workloads and the slow growth in the capacity of the main memory devices. We also observe that recent scientific applications are processing huge amount of data never been seen before. Scientific applications show a tendency to repeat operations on intermediate data using loops, but such access patterns are hard to be appropriately handled by LRU, which is generally used as a page replacement policy in operating systems. In this paper, we propose an adaptive page replacement (APR) scheme, which deals with looping access patterns in scientific applications. APR detects various looping access patterns online and handles them while considering the resulting performance. It can be implemented using limited events or information (e.g., page faults and access bits) provided by the virtual memory subsystem of operating systems. We evaluate APR through trace-driven simulation with traces extracted from workloads in the SPLASH-2x benchmark. Throughout the comparison with multiple previous schemes, we demonstrate that APR successfully improves the performance over the previous schemes by avoiding their downsides.