Re-Reference Interval Prediction Swap Policy in the Linux Kernel

Abstract

Modern computing systems are placing ever greater pressure on their memory management systems. The current means of managing the page cache in the Linux kernel is a binary ranking standard through which cached pages are stored either in an active list or an inactive list and managed by an approximation of a least recently used (LRU) algorithm. Recent endeavors in processor caching have revealed the opportunity for increased performance resulting from refining LRU memory management algorithms. I sought to determine the feasibility of replacing the current pseudo-LRU page cache system with one based on re-reference interval prediction (RRIP). This was achieved this by exploring the current Linux kernel to understand how exactly the page cache is managed, programming and implementing the custom RRIP page level replacement policy, and performing benchmark tests to determine the change in performance of the new system. Conducting this research determined the potential of RRIP in the Linux kernel’s page cache system to be at least on par with the current architecture and paved the way for future opportunities to expand and refine RRIP in the context of the page caching.