Towards a general framework to model, analyze and optimize real-time systems with GPUs

Abstract

Although the computational power and efficiency of GPUs would clearly benefit emerging real-time applications such as smart mobility, the adoption of such accelerators is being hindered by the poorly documented nature of their internal scheduling mechanisms. There is presently an intense research effort to propose solutions to enable the safe usage of GPUs in real-time applications [1] [2], although their applicability remains a challenge. In this extended abstract we present our methodology to safely incorporate GPUs into real-time systems. We propose a comprehensive framework that builds upon existing and validated tools and techniques, based on three main aspects: (1) an extension of an industry relevant meta-model, called MAST-2, to support GPUs, (2) leveraging time partitioning to control the access to the GPUs, which enables the application of existing WCRT (Worst-Case Response Time) analysis techniques, and (3) an optimization framework that takes advantage of the previous metamodel and analysis, to construct optimized time partitions.