Scheduling Tasks on Intermittently-Powered Real-Time Systems Public Deposited

Downloadable Content

Download PDF
  • Islam, Bashima
    • Affiliation: College of Arts and Sciences, Department of Computer Science
  • Batteryless systems go through sporadic power on and off phases due to intermittently available energy; thus, they are called intermittent systems. Unfortunately, this intermittence in power supply hinders the timely execution of tasks and limits such devices’ potential in certain application domains, e.g., healthcare, live-stock tracking. Unlike prior work on time-aware intermittent systems that focuses on timekeeping [1, 2, 3] and discarding expired data [4], this dissertation concentrates on finishing task execution on time. I leverage the data processing and control layer of batteryless systems by developing frameworks that (1) integrate energy harvesting and real-time systems, (2) rethink machine learning algorithms for an energy-aware imprecise task scheduling framework, (3) develop scheduling algorithms that, along with deciding what to compute, answers when to compute and when to harvest, and (4) utilize distributed systems that collaboratively emulate a persistently powered system. Scheduling Framework for Intermittently Powered Computing Systems. Batteryless systems rely on sporadically available harvestable energy. For example, kinetic-powered motion detector sensors on the impalas can only harvest energy when the impalas are moving, which cannot be ascertained in advance. This uncertainty poses a unique real-time scheduling problem where existing real-time algorithms fail due to the interruption in execution time. This dissertation proposes a unified scheduling framework that includes both harvesting and computing. Imprecise Deep Neural Network Inference in Deadline-Aware Intermittent Systems. This dissertation proposes Zygarde- an energy-aware and outcome-aware soft-real-time imprecise deep neural network (DNN) task scheduling framework for intermittent systems. Zygarde leverages the semantic diversity of input data and layer-dependent expressiveness of deep features and infers only the necessary DNN layers based on available time and energy. Zygarde proposes a novel technique to determine the imprecise boundary at the runtime by exploiting the clustering classifiers and specialized offline training of the DNNs to minimize the loss of accuracy due to partial execution. It also proposes a single metric, η to represent a system’s predictability that measures how close a harvesterâs harvesting pattern is to a constant energy source. Besides, Zygarde consists of a scheduling algorithm that takes available time, available energy, impreciseness, and the classifier's performance into account. Scheduling Mutually Exclusive Computing and Harvesting Tasks in Deadline-Aware Intermittent Systems. The lack of sufficient ambient energy to directly power the intermittent systems introduces mutually exclusive computing and charging cycles of intermittently powered systems. This introduces a challenging real-time scheduling problem where the existing real-time algorithms fail due to the lack of interruption in execution time. To address this, this dissertation proposes Celebi, which considers the dynamics of the available energy and schedules when to harvest and when to compute in batteryless systems. Using data-driven simulation and real-world experiments, this dissertation shows that Celebi significantly increases the number of tasks that complete execution before their deadline when power was only available intermittently. Persistent System Emulation with Distributed Intermittent System. Intermittently-powered sensing and computing systems go through sporadic power-on and off periods due to the uncertain availability of energy sources. Despite the recent efforts to advance time-sensitive intermittent systems, such systems fail to capture important target events when the energy is absent for a prolonged time. This event miss limits the potential usage of intermittent systems in fault- intolerant and safety-critical applications. To address this problem, this dissertation proposes Falinks, a framework that allows a swarm of distributed intermittently powered nodes to collaboratively imitate the sensing and computing capabilities of a persistently powered system. This framework provides power-on and off schedules for the swamp of intermittent nodes which has no communication capability with each other.
Date of publication
Resource type
Rights statement
  • In Copyright - Educational Use Permitted
  • Nirjon, Shahriar SN
  • Jeffay, Kevin KJ
  • Porter, Donald E. DEP
  • Jiang, Xiaofan XJ
  • Duggirala, Parasara Sridhar PSD
  • Doctor of Philosophy
Degree granting institution
  • University of North Carolina at Chapel Hill Graduate School
Graduation year
  • 2021

This work has no parents.