Application of Machine Learning Techniques on Prediction of Future Processor Performance

Today, processors utilize many datapath resources with various sizes. In this study, we focus on single thread microprocessors, and apply machine learning techniques to predict processors' future performance trend by collecting and processing processor statistics. This type of a performance prediction can be useful for many ongoing computer architecture research topics. Today, these studies mostly rely on history-and threshold-based prediction schemes, which collect statistics and decide on new resource configurations depending on the results of those threshold conditions at runtime. The proposed offline training-based machine learning methodology is an orthogonal technique, which may further improve the performance of such existing algorithms. We show that our neural network based prediction mechanism achieves around 70% accuracy for predicting performance trend (gain or loss in the near future) of applications. This is a noticeably better result compared to accuracy results obtained by naïve history based prediction models.