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 re-search. While much of this work has focused on single-GPU systems, multiple GPUs may be used for further ben-efits. Similar to CPUs in multiprocessor systems, GPUs in multi-GPU systems may be managed using partitioned, clus-tered, or global methods, independent of CPU organiza-tion. This gives rise to many combinations of CPU/GPU or-ganizational methods that, when combined with additional GPU management options, results in thousands of “rea-sonable ” configuration choices. In this paper, we explore real-time schedulability of several categories of configura-tions for multiprocessor, multi-GPU systems that are pos-sible under GPUSync, a recently proposed highly config-urable real-time GPU management framework. Our analy-sis includes the careful consideration of GPU-related over-heads. We show system configuration strongly affects real-time schedulability. We also identify which configurations of-fer the best schedulability in order to guide the implementa-tion of GPU-based real-time systems and future research.