•   When: Friday, April 08, 2016 from 02:00 PM to 04:00 PM
•   Speakers: Mohammad Atiqul Haque
•   Location: Nguyen Engineering, Room 4201
•   Export to iCal

Abstract

Energy conservation and reliability management are important design objectives for real-time embedded systems. The existing solutions that simultaneously consider energy and reliability dimensions are mostly developed for single-processor systems. On the other hand, real-time embedded systems are increasingly implemented on multiprocessor/multicore platforms. Multiprocessor systems, by naturally providing hardware redundancy and parallel execution ability, offer a wide design spectrum for reliability, energy, and timeliness dimensions. The main theme of this dissertation is to incorporate reliability management techniques in energy-efficient multiprocessor real-time systems.

This dissertation first addresses the problem of minimizing energy consumption on a dual-processor real-time standby-sparing system. The main copies of the tasks are executed on the first (primary) processor at low voltage/frequency levels. The backup copies are allocated to the second (spare) processor, and they are canceled when the main copies complete successfully. The dissertation proposes offline and online techniques to minimize the execution overlap between the main and backup copies, while tolerating transient faults and enhancing energy savings in a timely manner. Fixed-priority and dynamic-priority real-time systems are considered separately. Second, the dissertation presents a generalized reliability-oriented power management framework for multiprocessor real-time systems. The objective is to satisfy arbitrary task-level reliability levels with minimum energy consumption, by using task replication and voltage/frequency scaling. An offline optimization framework, enhanced by a suite of online adaptation techniques, is developed. Finally, the dissertation considers a recently proposed fault model in which the system is subject to a burst of transient faults during a time window of bounded length. During a fault burst, multiple tasks and even some recovery operations may be impacted. The dissertation derives exact schedulability conditions and proposes a recovery framework for real-time systems which may be under the influence of a fault burst at an arbitrary time.