Padel: Priority-Based Real-Time Scheduling for GPUs

Abstract

Graphics processing units (GPUs) have become increasingly prevalent in many platforms, including real-time systems, due to their massive architectural parallelism and significant performance. Power and energy management, reducing deadline miss rate (DMR), and efficient allocation of resources to tasks are important design challenges in exploiting GPUs within real-time platforms. One of the most considerable challenges in designing firm real-time systems is allocating GPU resources to tasks in a way that enables as many tasks as possible to be completed correctly within their deadlines while minimizing energy consumption. However, since multiple tasks can be handled concurrently across many cores, power consumption poses a serious limitation. In this work, we introduce Padel, a real-time GPU scheduler that employs spatial multitasking to enhance the utilization of resources, performance, and energy efficiency in GPUs. Our experimental results reveal that when the system is overloaded, Padel can reduce the DMR by 23%, and energy consumption by 8% compared to the state of the art.