DACA: a feedback control-based approach to approximate computing
Open Access
- Author:
- Jang, Oh Young
- Graduate Program:
- Computer Science and Engineering
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- July 27, 2012
- Committee Members:
- Mahmut Taylan Kandemir, Thesis Advisor/Co-Advisor
Chitaranjan Das, Thesis Advisor/Co-Advisor - Keywords:
- approximate computing
control theory
bodytrack - Abstract:
- Inaccuracy in computation has usually been considered with a negative connotation, and therefore, conventional computing systems have always been designed with a strict notion of correctness. However, inaccuracy or approximation is not always bad since several application domains are intrinsically tolerant to varying degrees of relaxation in accuracy, and thus, such a property can be exploited for significant gains in application performance or fault-tolerance. This concept, known as approximate computing or soft computing, has been recently applied to several application domains primarily for quantifying the tradeoffs between potential performance gains and accuracy losses. A weakness of all these studies is that the error bound is not strictly limited. The main contribution of this work is a feedback control based scheme, called DACA, to approximate computing that can be used to constrain the loss in accuracy, while maximizing potential performance benefits. We demonstarate the design of the feeback controller to dynamically actuate the approximate computing mecahanism using an object tracking application, Bodytrack. In addition, we also propose a roll-back technique to control large variance in inaccuracy that cannot be easily controlled by the feedback mechanism. Our evaluations indicate that the proposed approach can improve performance by as much as 260%, while guaranteeing inaccuracy limit between 10% to 30%.