Evaluating a Machine Learning-based Approach for Cache Configuration

As the systems perform progressively complex tasks, the search for energy efficiency in computational systems is constantly increasing. The cache memory has a fundamental role in this issue. Through dynamic cache reconfiguration techniques, it is possible to obtain an optimal cache configuration that minimizes the impacts of energy losses. To achieve this goal, a precise selection of cache parameters plays a fundamental role. In this work, a machine learning-based approach is evaluated to predict the optimal cache configuration for different applications considering their dynamic instructions and a variety of cache parameters, followed by experiments showing that using a smaller set of application instructions it is already possible to obtain good classification results from the proposed model. The results show that the model obtains an accuracy of 96.19% using the complete set of RISC-V instructions and 96.33% accuracy using the memory instructions set, a more concise set of instructions that directly affect the cache power model, besides decreasing the model complexity.