Adapted discrete-based entropy cache replacement algorithm

It is fundamental to predict cache line reuse in order to improve miss rates and consequently performance. In order to estimate cache line reuse, we assume that a sequence of cache line addresses can be treated randomly as characters in an arbitrary language such as proposed by the Information Theory. In this theory, the occurrence of each character previously occurred in a message composed by a sequence of characters can be used as a measurement of information - named entropy. The computation of the entropy of a program working set helps to estimate the chances of line reuse. We propose Adapted Discrete-based Entropy Algorithm (ADEA), a novel cache line replacement inspired by the Information Entropy which measures the discrete entropy by capturing the referential locality of programs by estimating the chances of cache line reuse. Furthermore, ADEA presents functions that control the essency and recency of memory access to avoid cache pinning. We present ADEA circuit complexity and show that it is comparable to other implementations. By modeling and evaluating ADEA using Simple Scalar for an L2 cache of an OOO processor, results show that ADEA´s miss rate is up to 60% lower than LRU and 36% compared to adaptive policies Least Recently Used Insertion Policy (LIP) and Bimodal Insertion Policy (BIP). Furthermore, the use of ADEA´s decay functions improves its miss rates up to 46% by avoiding pinning.