A page replacement algorithm based on frequency derived from reference history

Page replacement algorithm is one of the core components in modern operating systems. It decides which victim page to evict from main memory by analyzing attributes of pages referenced. The evicted page is then moved to backing store in the memory hierarchy, and moved back to main memory once referenced again. The technique that utilizes storage as part of memory is called swapping. However, there is a non-trivial performance gap between memory and storage. For example, performance of permanent storage like Solid-State Disk (SSD) is much slower, e.g. 10⁴ longer write latency, than DRAM [9]. As a result, swapping between main memory and storage causes system performance to a discernible drop. Nevertheless, a higher hit ratio of page replacement algorithm implies less I/O waits to storage, and consequently a better performance overall.

In this paper we propose a log-based page replacement algorithm that assumes better hints for page replacement can be approached through analysis of page reference history. The algorithm selects victim page that holds lowest reference rate in a window-sized log. A simulation shows that our method outperforms conventional page replacement algorithms by 11+ at best.