EDZL Scheduling and Schedulability Analysis for Performance Asymmetric Multiprocessors

Abstract

Due to the rapid development in hardware design, multiprocessors are widely used in modern industry, especially heterogeneous multiprocessors. Heterogeneous multiprocessor architectures allow embedded real-time systems to better match computing resources to each application's needs and dynamic workload requirements, thereby providing many opportunities for improved performance with reduced power consumption. Unfortunately, guaranteeing real-time requirements on heterogeneous multiprocessors remains a critical problem due to the lack of appropriate scheduling algorithms and analysis methods. “Dhall’s effect” is a problem which arises when classic scheduling algorithms of single processor are applied to multiprocessors, such as EDF and RM. They fail to schedule some low-utilization task sets even though many processors are free. EDZL (Earliest Deadline First until Zero-Laxity) scheduling can solve “Dhall’s effect” problem for performance asymmetric multiprocessors. In this work, we consider EDZL scheduling for performance asymmetric multiprocessor scheduling. EDZL scheduling is a hybrid scheduling policy that combines Earliest Deadline First (EDF) and Least Laxity First (LLF) scheduling. In EDZL scheduling, tasks with zero laxity are given the highest priority while other tasks are ranked by deadlines. EDZL scheduling has been shown to outperform other scheduling policies such as global EDF scheduling on identical multiprocessors. In the present work, we show that EDZL scheduling is also effective on performance asymmetric multiprocessors, and present an efficient EDZL schedulability test. We have implemented several schedulability tests including our EDZL test and compare their performance through extensive simulations. Our experimental results show that EDZL scheduling is able to schedule up to 20% more task sets than global EDF. We also show that EDZL schedulability test outperforms global RM schedulability test for performance asymmetric multiprocessors too. Moreover, our new EDZL schedulability test can accept up to 30% more schedulable task sets than a presently exiting one.