Access Type

Open Access Dissertation

Date of Award

January 2015

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Computer Science

First Advisor

Nathan Fisher

Abstract

We introduce a new design metric called system-resiliency which characterizes the maximum unpredictable

external stresses that any hard-real-time performance mode can withstand. Our proposed systemresiliency

framework addresses resiliency determination for real-time systems with physical and hardware

limitations. Furthermore, our framework advises the system designer about the feasible trade-offs between

external system resources for the system operating modes on a real-time system that operates in a

multi-parametric resiliency environment.

Modern multi-modal real-time systems degrade the system’s operational modes as a response to unpredictable

external stimuli. During these mode transitions, real-time systems should demonstrate a reliable

and graceful degradation of service. Many control-theoretic-based system design approaches exist. Although

they permit real-time systems to operate under various physical constraints, none of them allows

the system designer to predict the system-resiliency over multi-constrained operating environment. Our

framework fills this gap; the proposed framework consists of two components: the design-phase and runtime

control. With the design-phase analysis, the designer predicts the behavior of the real-time system for

variable external conditions. Also, the runtime controller navigates the system to the best desired target

using advanced control-theoretic techniques. Further, our framework addresses the system resiliency of

both uniprocessor and multicore processor systems.

As a proof of concept, we first introduce a design metric called thermal-resiliency, which characterizes

the maximum external thermal stress that any hard-real-time performance mode can withstand. We verify

the thermal-resiliency for the external thermal stresses on a uniprocessor system through a physical testbed.

We show how to solve some of the issues and challenges of designing predictable real-time systems that

guarantee hard deadlines even under transitions between modes in an unpredictable thermal environment

where environmental temperature may dynamically change using our new metric.

We extend the derivation of thermal-resiliency to multicore systems and determine the limitations of

external thermal stress that any hard-real-time performance mode can withstand. Our control-theoretic

framework allows the system designer to allocate asymmetric processing resources upon a multicore proiii

cessor and still maintain thermal constraints.

In addition, we develop real-time-scheduling sub-components that are necessary to fully implement our

framework; toward this goal, we investigate the potential utility of parallelization for meeting real-time

constraints and minimizing energy. Under malleable gang scheduling of implicit-deadline sporadic tasks

upon multiprocessors, we show the non-necessity of dynamic voltage/frequency regarding optimality of

our scheduling problem. We adapt the canonical schedule for DVFS multiprocessor platforms and propose

a polynomial-time optimal processor/frequency-selection algorithm.

Finally, we verify the correctness of our framework through multiple measurable physical and hardware

constraints and complete our work on developing a generalized framework.

Download

Share

COinS