A categorical study on cache replacement policies for hierarchical cache memory

Abstract

Cache memory plays an important role in the in-memory computation in memory-intensive applications. Hierarchical cache design is used to increase the capacity of cache to handle large working set. The last level cache (LLC) does not strictly follow the temporal locality of program, so it becomes challenging to identify the blocks that will not be reused (dead block). In this paper, we have performed a detail survey on different techniques to detect the dead blocks early in the cache memory and improve the hit rate of cache replacement algorithm. Belady’s optimal solution detects the dead block by analyzing the future of blocks, which is completely un-realistic. Many researches have been done to detect dead block practically by observing the previous access pattern. Many algorithms are proposed to improve the performance of traditional replacement policies by considering different additional information. Most of the algorithm aims to reduce the miss count by retaining the blocks that will be reused before eviction (live blocks). Recent study observes that the cost of all the cache miss are not uniform in nature. So, some researchers have distinguished between high-cost block and low-cost block. The overall cost can be reduced by retaining the high-cost block in memory, with little higher miss count. It is observed that by managing cache miss un-coordinately among the different levels of cache memory, it is not possible to obtain maximum utilization of memory. Many adaptive algorithms have been proposed to maintain balance between the over-utilized blocks and underutilized blocks by the displacement of blocks. In this survey, we have categorized the practically implemented techniques into different classes based on their basic principle of cache replacement.