Downloadable Content

Download PDF
Last Modified
  • March 21, 2019
  • Yang, Kecheng
    • Affiliation: College of Arts and Sciences, Department of Computer Science
  • Real-time scheduling analysis is crucial for time-critical systems, in which provable timing guarantees are more important than observed raw performance. Techniques for real-time scheduling analysis initially targeted uniprocessor platforms but have since evolved to encompass multiprocessor platforms. However, work directed at multiprocessors has largely focused on symmetric platforms, in which every processor is identical. Today, it is common for a multiprocessor to include heterogeneous processing elements, as this offers advantages with respect to size, weight, and power (SWaP) limitations. As a result, realizing modern real-time systems on asymmetric multiprocessor platforms is an inevitable trend. Unfortunately, principles and mechanisms regarding real-time scheduling on such platforms are relatively lacking. The goal of this dissertation is to enrich such principles and mechanisms, by bridging existing analysis for symmetric multiprocessor platforms to asymmetric ones and by developing new techniques that are unique for asymmetric multiprocessor platforms. The specific contributions are threefold. First, for a platform consisting of processors that differ with respect to processing speeds only, this dissertation shows that the preemptive global earliest-deadline-first (G-EDF) scheduler is optimal for scheduling soft real-time (SRT) task systems. Furthermore, it shows that semi-partitioned scheduling, which is a hybrid of conventional global and partitioned scheduling approaches, can be applied to optimally schedule both hard real-time (HRT) and SRT task systems. Second, on platforms that consist of processors with different functionalities, tasks that belong to different functionalities may process the same source data consecutively and therefore have producer/consumer relationships among them, which are represented by directed acyclic graphs (DAGs). End-to-end response-time bounds for such DAGs are derived in this dissertation under a G-EDF-based scheduling approach, and it is shown that such bounds can be improved by a linear-programming-based deadline-setting technique. Third, processor virtualization can lead a symmetric physical platform to be asymmetric. In fact, for a designated virtual-platform capacity, there exist an infinite number of allocation schemes for virtual processors and a choice must be made. In this dissertation, a particular asymmetric virtual-processor allocation scheme, called minimum-parallelism (MP) form, is shown to dominate all other schemes including symmetric ones.
Date of publication
Resource type
Rights statement
  • In Copyright
  • Anderson, James
  • Smith, Donelson
  • Nirjon, Shahriar
  • Bini, Enrico
  • Baruah, Sanjoy
  • Doctor of Philosophy
Degree granting institution
  • University of North Carolina at Chapel Hill Graduate School
Graduation year
  • 2018

This work has no parents.