Access Type

Open Access Dissertation

Date of Award

January 2019

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Computer Science

First Advisor

Nathan . Fisher

Abstract

For hard real-time systems, schedulability of a task set is paramount. If a task set is not deemed schedulable under all conditions, the system may fail during operation and cannot be deployed in a high risk environment. Schedulability testing has typically been separated from worst-case execution time (WCET) analysis. Each task’s WCET value is calculated independently and provided as input to a schedulability test. However, a task’s WCET value is influenced by scheduling decisions and the impact of cache memory. Thus, schedulability tests have been augmented to include cache-related preemption delay (CRPD). From this classical perspective, the effect of cache memory on WCET and schedulability is always negative; increasing execution times and demand. In this work we propose a new positive perspective, where cache memory benefits multi-threaded tasks by scheduling threads in a manner that shares values predictably.

This positive perspective is reached by integrating, rather than separating the disciplines of schedulability analysis and worst-case execution time. These integrated techniques are referred to as the BUNDLE family of worst-case execution time and cache overhead

(WCETO) analysis and scheduling algorithm. WCETO calculation divides the task’s structure into conflict free regions and calculates a bound utilizing explicit understanding of the thread-level scheduling algorithm. Conflict free regions are utilized by the scheduling algorithm, which associates with each region a thread container called a bundle. At any time only one bundle may be active, and only threads of the active bundle may execute on the processor.

The BUNDLE family of scheduling algorithms developed in this work increase in scope from BUNDLE through ITCB-DAG. As the fundamental contribution, BUNDLE and BUNDLEP apply to a single multi-threaded task running on a uniprocessor architecture with a single level direct mapped instruction cache. NPM-BUNDLE expands the positive perspective to multiple tasks on a uniprocessor system. With ITCB-DAG bringing BUNDLE’s analysis and scheduling techniques to multi-processor systems.

Each of the scheduling algorithms require a novel hardware mechanism to anticipate execution and make scheduling decisions. To support anticipation of execution, a novel XFLICT interrupt is proposed. It is a simple mechanism that emulates the behavior of hardware breakpoints. An implementation of the BUNDLEP analytical techniques, scheduling algorithm, and XFLICT interrupt is available as a simulated platform for further research and extension.

Future work is planned to expand BUNDLE’s positive perspective and increase adoption. The most significant barrier to adoption is the ability to deploy BUNDLE’s scheduling algorithm, this mandates a viable and available hardware or software mechanism to anticipate

execution. NPM-BUNDLE is limited to non-preemptive multi-task scheduling and analysis, support for preemptive scheduling will increase the positive impact of BUNDLE’s integrated perspective.

Share

COinS