Hybrid Machine Learning Model for Cache Performance Prediction and Design Space Exploration

This work proposes a hybrid machine-learning model to predict a level 1 data (L1D) cache miss rate value of diversified applications with different cache configurations and application characteristics. The proposed model quickly predicts a miss rate of the L1D cache. Moreover, the proposed model can simultaneously predict the miss rate value of a new(unseen) application for twenty-eight cache configurations as precisely as possible without extra simulations. Thus, it quickly provides a miss rate value faster than a simulator. Additionally, we could perform an application’s cache design space exploration by getting predicted miss rate values from the proposed model, like the simulator. The proposed model’s mean absolute and root mean squared errors are 0.6% and 1.1% by ten-fold cross-validation. Speed up was achieved 12x to 148x compared to the architectural simulator for a miss rate of twenty-eight cache configurations.