Using low-power platforms for Evolutionary Multi-Objective Optimization algorithms

Nowadays, the application of Evolutionary Multi-Objective Optimization (EMO) algorithms in real-time systems receives considerable interest. In this context, the energy efficiency of computational systems is of paramount relevance. Recently, the use of embedded systems based on heterogeneous (CPU + GPU) platforms is consistently increasing. For example, NVIDIA Jetson cards are low-power computers designed for development of embedded applications. They incorporate Tegra processors which feature a CUDA-capable GPU. This way, Jetson cards can be considered as a prototype of low-power computer of High-Performance Computing. In this work, our interest is focused on the NSGA-II algorithm, a well-known representative of EMO algorithms. The strength of NSGA-II lies in its Non-Dominated Sorting (NDS) procedure of a population of individuals. Our purpose on the low-power computers is twofold: to define and evaluate the parallel NSGA-II versions with major focus on NDS procedure on the Jetson platforms and to determinate the size of NSGA-II problems which can be solved. The results show that the parallel version which achieves the best performance depends on the objectives functions and the frequencies of the clocks of the cores and memory of the GPU. The analysis of the results shows the capability of the Jetson as a low-consumption platform which allows to accelerate the execution of instances of the state-of-the-art EMO algorithm—NSGA-II.