Résumé
Heterogeneous multicore systems, such as ARM big.LITTLE, use different types of processors to conciliate high performance with low energy consumption. A question that concerns such systems is how to find the best hardware configuration (type and frequency of processors) for a program. Current solutions are either completely dynamic, based on in-vivo profiling, or completely static, based on supervised machine learning. Whereas the former approach can bring unwanted runtime overhead, the latter fails to account for diversity in program inputs. In this paper, we design and evaluate a compilation strategy, JINN-C, that perform statistical regression on function arguments, so as to match parameters with ideal hardware configurations at runtime. We show that JINN-C, implemented in the Soot compiler, can predict the best configuration for a suite of Java and Scala programs running on an Odroid XU4 board, while outperforming prior techniques such as ARM's GTS and CHOAMP, a recently released static program scheduler.