Exploring the Multitude of Real-Time Multi-GPU Configurations

Motivated by computational capacity and power efficiency, techniques for integrating graphics processing units (GPUs) into real-time systems have become an active area of research. While much of this work has focused on single-GPU systems, multiple GPUs may be used for further benefits. Similar to CPUs in multiprocessor systems, GPUs in multi-GPU systems may be managed using partitioned, clustered, or global methods, independent of CPU organization. This gives rise to many combinations of CPU/GPU organizational methods that, when combined with additional GPU management options, results in thousands of "reasonable" configuration choices. In this paper, we explore real-time schedulability of several categories of configurations for multiprocessor, multi-GPU systems that are possible under GPUSync, a recently proposed highly configurable real-time GPU management framework. Our analysis includes the careful consideration of GPU-related overheads. We show system configuration strongly affects real time schedulability. We also identify which configurations offer the best schedulability in order to guide the implementation of GPU-based real-time systems and future research.