GPU: A New Enabling Platform for Real-Time Optimization in Wireless Networks

Abstract

Optimization methods are a common tool to maximize the performance of wireless networks and systems. When addressing complex optimization problems in wireless networks, a key technical challenge is to find an optimal or near-optimal solution in real-time, especially when such a timing constraint is extremely short. Due to this challenge, there is usually a serious disparity between what a system can achieve optimally (if an optimal solution were found in real time) and what is actually achieved in the field (due to the use of fast heuristics). in this article, we present a novel approach that exploits problem decomposition techniques and the massive parallel processing capability of GPU platforms to address this challenge. Under the new approach, an original complex optimization problem is first decomposed into a large number of small and mutually independent sub-problems. Then the resulting sub-problems are fitted into massively parallel GPU cores and solved simultaneously. The optimal (or near-optimal) solution is chosen among the solutions from all the parallel sub-problems solved by GPU. We use the classic proportional-fair (PF) scheduling problem in 5G cellular networks as a case study to illustrate this approach. Finally, we briefly review recent advances in applying this approach to addressing a wide array of real-time optimization problems in wireless networks.