Inaccuracy in computation has usually been considered with a negative connotation and, therefore, conventional computing systems have 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 gain in application performance or fault-tolerance. The motivation of this EAGER project is to investigate the feasibility of utilizing such approximation, also known as "soft computing", for data-intensive applications for predicting the performance-power-accuracy trade-offs. The research consists of three intertwined tasks. The first task would examine a variety of high performance computing (HPC) and MapReduce style data analytic applications, and determine which classes of applications are suitable for soft computing. The second component of the research is aimed at developing appropriate techniques for facilitating soft computing, while the last task focuses on examining the possibility of developing a control theoretic model for formalizing the various tradeoff analysis.
This project aims at demonstrating that it is possible to achieve significant power and performance gain for a wide variety of data intensive applications through soft computing. The approach adopted in this research has the potential to influence the programming paradigm for many classes of scientific and business applications for optimizing the power-performance behavior. The cross-cutting nature of this research has potential to foster new research directions in several areas, spanning high performance computing, computer architecture, compilers, and system/application software. Undergraduate and graduate students involved in this research will get versatile training in several areas. The software tools developed in this research will be used in teaching existing and new courses, and will be made publicly available.
The focus of this EAGER project is to investigate the feasibility of utilizing approximation in computation, also known as soft computing, for understanding and predicting performance-power-accuracy trade-offs in several key application domains. The rationale for using soft computing is that many applications can intrinsically tolerate varying degree of error and thus, approximate computing can be used effectively in such cases to expedite computation and save energy. Since this is a new area of research, issues such as which applications can benefit from approximate computing, what are feasible techniques for employing approximate computing and what are the performance-power-accuracy tradeoffs are little explored. In this exploratory 2-year project, we have attempted to understand some of these issues along three dimensions. First, we have examined a variety of high performance computing (HPC), MapReduce style data analytic and graphic applications, and determined which classes of applications are suitable for soft computing. Second, we have developed several techniques for facilitating soft computing. Third, we have developed a formal method (control theoretic model) for conducting various tradeoff analyses. Inaccuracy in computation has usually been considered with a negative connotation. But, in this project we make a strong case for designing systems with acceptable inaccuracy for gains in performance and energy. Specifically, we found approximate computing suitable for applications from the following domains: (i) Weather prediction; (ii) Machine Learning; (iii) Data clustering; (iv) Video processing; and (v) Mobile platforms. Energy-efficient computing is a critical area of research that would benefit the society in several ways. Based on our interaction with several researchers in the hardware, software, algorithm and application areas, we expect that approximate computing will receive increasing attention in coming years. This EAGER project is an initial attempt in this open area of research.
