Building A Game Benchmark for Cooperative CPU-GPU with Pseudo User-interaction

Power and performance optimization on mobile platforms has become a significant research topic because of the power consumption and battery life problem. However, it is difficult to evaluate the effectiveness of an optimization due to the interactive nature of mobile platforms. Thus, there is a growing need for benchmarks to help researchers overcome the user interference such as interaction styles and user interests. This paper introduces and characterizes PUBench, a game benchmark for cooperative CPU-GPU with Pseudo User-interaction, which is designed to cover a broad range of interactive nature, workload nature, and power nature of mobile games. To facilitate the development of PUBench, we first propose a Factor Profiling Architecture (FPA) that analyzes related factors and the relations among different factors from top to bottom for a mobile platform. By FPA, we draw a conclusion that the de facto factors should be covered in PUBench are the APP-OS interaction and the CPU-GPU interaction. We apply a queuing model to further quantify these factors. Then a PUBench prototype which adopts the user-interaction as the first-class citizen is outlined and designed based on the quantification. We implement the prototype as an Android application and show that PUBench can accurately imitate the interactive game behaviors in both workload and power/performance by experiments. Finally, we show a case study of PUBench by applying it to evaluate the power management policies on the Odroid-XU3 platform.