Scheduling Processing Graphs of Gang Tasks on Heterogeneous Platforms

Artificial-intelligence-powered real-time systems typically consist of numerous gang tasks, such as computations on graphics processing units (GPUs), that are interconnected by data-flow dependencies. Despite their relevance in many applications, scheduling processing graphs of gang tasks has received limited attention. This paper presents scheduling techniques and response-time analysis for such systems on heterogeneous computing platforms. Response-time bounds of a graph of gang tasks are presented when scheduled under a work-conserving or semi-work-conserving scheduler. Techniques to support multiple graphs using federated scheduling techniques are also presented. Experimental evaluations and a case study on a computer vision application are presented to demonstrate the effectiveness of the proposed approach.