Dynamic pseudo-LRU for balancing disturbance

Abstract

Cache is a basic computer architecture for better performance in modern computer. Though small size of cache which cannot have all blocks in memory, blocks in cache should be replaced with new blocks from memory. In this kind of view, many cache replacement policies come into the world. The most well-known policy is Least Recently Used (LRU) policy. LRU gives positions to all blocks, moves new or re-referenced block to first position and evicts the last position's block. Tree-based PseudoLRU (PLRU) is similar policy with LRU which uses binary tree to express position. PLRU has better space efficiency than LRU. However, this tree-based system makes unnecessary disturbances to non-referenced blocks in a set. Therefore, we make a Counter-based Pseudo-LRU (CpLRU) which diminishes disturbance of non-referenced blocks. Finally, we propose the Dynamic Pseudo-LRU (DpLRU) which monitors CpLRU and PLRU and chooses better policy. To make Counter-based Pseudo-LRU, we analyze three methods that could be used on PLRU. The first one is replacement position of new incoming block. Second one is whether to change a bit when there was hit in second quarter of a set. Last is whether to maintain the counter when there was change in counting part. This CpLRU improves MPKI in single thread by 0.07MPKI than LRU and PLRU. Dynamic Pseudo-LRU is made by applying Set Dueling between PLRU and CpLRU. DpLRU shows 0.4MPKI less than PLRU.