Applying machine learning to enhance the cache performance using reuse distance

Abstract

The motivation for the search of efficient cache management policies is due to the drawback of the large number of CPU cycles taken to fetch a block from the main memory to the last level cache. Recently, Researchers are using most recent advancements in machine learning (ML) to optimise the performance of cache. They designed different models using ML techniques to improve the performance of cache. Unfortunately, the schemes suffer from several drawbacks like storage complexity incurred while storing the extra bits required for prediction, excess time required for prediction, increase in difficulty of reuse prediction when workloads are increased, increase in energy consumption during the prediction process and lack in achieving expected accuracy. In order to overcome these drawbacks, cache reuse prediction is analysed using classification problem, which is actual view of the learning problem. The vital parameters used are the Frequency of access and reuse distance, in which the latter is modelled as stack distance. As an innovative idea, reuse distance is calculated based on status of each block into account. In addition, numerical nature of the dataset used in this article, helps in exploration of the ensemble-based ML, which is less frequently experimented into the context of cache memories rather than the deep learning approach. By applying suitably placed models, the memory references with high reuse capabilities are fetched into the cache. They are predicted with the aid of classification and cross-validation on the feature sets, which improve the accuracy and sensitivity by 6.2% and 5.46% respectively, when compared to the existing works.