Effective feature set construction for SVM-based hot method prediction and optimisation

Abstract

Optimising compilers rely on profiling to identify the target regions for optimising the input programme. Although profiling is accurate, it incurs a lot of overhead, an obstacle to achieving considerable performance improvement. Alternatively, machine learning-based offline prediction of hot methods that form vital target segments, is bound to eliminate the runtime overhead. In this work, we develop and implement support vector machines-based hot method prediction models trained on effective static programme features generated by a new 'knock-out' algorithm. When trained using low level virtual machine (LLVM) environment, it is possible to predict the frequently called and the long running hot methods with 61% and 68% accuracy. Selective optimisation of the predicted hot methods before programme execution provides substantial performance improvement over default LLVM optimisation.