Minseon Cho  Donghyun Kang

10.5626/JOK.2021.48.9.981

cache replacement policy Multi-layer perceptron lock-unlock LRU

The LRU (least-recently used) page replacement policy has been designed to enhance the cache hit ratio by moving the page that is repeatedly accessed on the cache, to the head of the list. However, the LRU policy sometimes incurs a situation of system stall (or wait) because it requires lock-unlock commands to move each page. In this paper, we propose a new page replacement policy, called L2LRU(Learning-based Lock-free LRU), that determines whether to move or not a page by learning the reuse distance of the page with deep-learning techniques. Unlike LRU, L2LRU moves the page to the position with a high possibility of access in the near future. For evaluation, we implemented L2LRU based on trace-driven simulation and used Microsoft Research Cambridge Trace as the input of the simulation. The results clearly confirmed that L2LRU reduced the number of lock-unlock commands by up to 91% compared to the traditional LRU policy.