Machine Learning for Cache Management 1.0


Giri Narasimhan

Product Owner(s):

Giri Narasimhan


Masoud Sadjadi


Machine Learning for Cache Management 1.0 purpose is to create a Cache Replacement Algorithm capable of having better hit ratio than experts such as ARC, LRU, CLOCK, or LFU by using Machine Learning techniques. As a result of 7 Sprints the team proposes a framework called LeCaR that using regret minimization is able to outperform ARC (strongest adversary) when cache available size is small relative to the size of working set. The first steps were to create the tools needed to analyze the learning rate of a ML algorithm and contrast the results with its competitors. Synthetic Traces is a feature that allows researchers to study the behavior of the weights and the hit ratio of any learning algorithm. The Pollution Counter feature was useful in order to understand the relationship between LeCaR and LFU. To make it appealing and user friendly, the Experiment Helper Windows application was developed.

Team Members

Wendy Aleman