Dynamic Analyses of Result Quality in Energy-Aware Approximate Programs.
详细信息
- 作者：RIngenburg ; Michael F.
- 学历：Ph.D.
- 年：2014
- 毕业院校：University of Washington
- Department：Computer Science and Engineering
- ISBN：9781303865381
- CBH：3618319
- Country：USA
- 语种：English
- FileSize：514514
- Pages：135
文摘
Energy efficiency is a key concern in the design of modern computer systems. One promising approach to energy-efficient computation,approximate computing,trades off output precision for energy efficiency. However,this tradeoff can have unexpected effects on computation quality. This thesis presents dynamic analysis tools to study,debug,and monitor the quality and energy efficiency of approximate computations. We propose three styles of tools: prototyping tools that allow developers to experiment with approximation in their applications,online tools that instrument code to determine the key sources of error,and online tools that monitor the quality of deployed applications in real time. Our prototyping tool is based on an extension to the functional language OCaml. We add approximation constructs to the language,an approximation simulator to the runtime,and profiling and auto-tuning tools for studying and experimenting with energy-quality tradeoffs. We also present two online debugging tools and three online monitoring tools. The first online tool identifies correlations between output quality and the total number of executions of,and errors in,individual approximate operations. The second tracks the number of approximate operations that flow into a particular value. Our online tools comprise three low-cost approaches to dynamic quality monitoring. They are designed to monitor quality in deployed applications without spending more energy than is saved by approximation. Online monitors can be used to perform real time adjustments to energy usage in order to meet specific quality goals. We present prototype implementations of all of these tools and describe their usage with several applications. Our prototyping,profiling,and autotuning tools allow us to experiment with approximation strategies and identify new strategies,our online tools succeed in providing new insights into the effects of approximation on output quality,and our monitors succeed in controlling output quality while still maintaining significant energy efficiency gains.