Category-wise Feature Analysis in SVM-based Hot Method Prediction

Abstract

The optimizing compiler makes critical decisions regarding the type of optimization to be performed and the target sequence on the input program. Program segments called hot methods that are frequently executed form important targets for optimization. Machine learnt prediction of these target segments for optimization relieves the virtual machine from profiling and the associated overhead. Because the efficiency of the prediction approaches based on machine learning depends on the core issue of feature selection, this study focuses on the impact of feature categories on the performance of the Support Vector Machine-based hot method predictive model. The standard sequential forward selection algorithm is used in the category-wise analysis and identification of the most influential feature category and the combination of categories. It is seen that the category combinations have larger impact on the performance of the predictive model than the individual feature categories and that the model is able to obtain the highest prediction accuracy of 67% when trained with a combination of the categories of method size and control flow graph.

Compared with default optimization heuristics enabled by Low Level Virtual Machine, the predictive model trained with method size and control flow graph achieves a modest speedup gain of 4% and 1% respectively for inlining and intra-procedural optimizations like constant propagation and loop unrolling on the MediaBench programs whereas on the UTDSP the results are mixed. Overall, the approach has the advantage of speeding up program execution in virtual machines.