The project aims at studying the fundamental question: how can use make modern parallel, heterogeneous, high-performance, embedded computing platforms in a way that delivers predictable and high performance. We focus on coarse grained predictability, meaning that we focus on subsystem level and that each subsystem should experience a predictable performance. In PREPPES we assume that the computing platform is inherently unpredictable and by using software techniques, such as server-based scheduling and hardware partitioning we will establish a level of predictability suitable of embedded systems.
The motivation for this project stems from the observation that more and more safety critical embedded systems are becoming increasingly reliant on high-performance computing platforms. For instance, autonomous machines such as vehicles and robots need enhanced perception capabilities though high data-rate sensors like stereo cameras, radars, lidars, etc. Contemporary sensors easily produce several hundreds of gigabytes of information per second and we anticipate that next generation sensors will produce data in the terabyte range.
This project will provide fundamental understanding of the practical boundaries of predictability of modern heterogeneous parallel computing platforms. While not striving for solutions that are considered as hard real-time systems, the aim is to create a better understanding of predictability and sources of unpredictability in in modern computing platforms.