ML-LIRS: Leveraging Machine Learning to Improve the LIRS Replacement Algorithm
Robert Fabbro, Zhong Chen, Song Jiang · 2021
While the LIRS replacement algorithm is more capable at exploiting locality of block accesses than LRU by using its inter-reference recency (IRR) locality measure, it may still make mistakes in its decision-making thereby mis-evicting blocks from the cache. By leveraging machine learning techniques, LIRS can be improved to make more accurate decisions in situations where the IRR is untrustworthy. The use of machine learning allows for the algorithm to be more understanding of a changing access pattern to improve prediction accuracy. This leads to a lower miss ratio over the life of a workload as the algorithm will be able to pick up both longer-term and short-term patterns that a static algorithm cannot.