Machine learning based approaches to predict and optimize hot methods

Abstract

Just-in-time compilers in a virtual machine environment generate machine code for a method in a program, when the method is first called for execution, incurring a code generation overhead. Dynamic optimizations overcome this limitation by optimizing the generated code during runtime with the profile collected from the current run of the program. This research work aims at developing a novel selective optimization technique, to serve as an effective alternative to the existing profile-based selective optimization of the input program. The technique uses the machine learning algorithm, the Support Vector Machine (SVM) in the construction of a heuristic, to predict the target sequences in a program, called hot methods for selective optimization. Two independent predictive models, one for predicting the frequently called, and another for the long running hot methods in a program, are developed. The new knock-out algorithm developed to perform feature reduction is an extension of the classical sequential backward elimination process, coupled with a new knock-out scheme to systematically eliminate one or more features at every iteration. The results obtained on the program execution time are compared against Low Level Virtual Machine s (LLVM) default optimization heuristics. The performance of the relearning predictive models on different combinations of SPEC, UTDSP and Mediabench benchmarks is very impressive, with an improvement of 10% and 21% respectively for the frequently called, and the long running hot methods, over the systems without relearning. The results of relearning confirm the effective predictability of hot methods by the machine learning based models. All these observations indicate that the technique of machine learning based hot method prediction is an effective selective optimization option, in any embedded system application running a virtual machine. newline newline newline