Predicting Efficient Execution with Source Code Analysis in a Heterogeneous Environment

Abstract

Finding a good schedule for the tasks of an application is a critical step for the efficient usage of heterogeneous systems. A good schedule can only be found with information about the tasks to be scheduled. In a dynamic system, this information is normally only available after each task is at least executed once, thereby creating an initial overhead until a good schedule can be created. Therefore, we introduce a method based on static code analysis and machine learning algorithms to predict the fastest processor of a given OpenCL task before runtime by classification which helps to reduce this initial overhead. We show how we used a static code analysis implementation based on Clang to generate training data on a set of 10 different heterogeneous processors including Intel, AMD and Nvidia GPUs, a Intel Xeon Phi and Intel CPUs. This training data was used to generate prediction models via several different machine learning algorithms including Random Forest and k-Nearest Neighbour and then evaluate the models by predicting the fastest processor out of two and more processors via classification.