| Scientific applications generally demand high performance as they are usually designed to solve time-critical large scale problems. Existing researches investigated optimizations for scientific applications to achieve good performance on modern microprocessors by carefully orchestrating the application behaviors to use the hardware resources efficiently. These techniques, however, require sophisticated compiler support to be applied on scientific applications to attain high performance. However, the impact of these optimizations on power consumption is not well-understood. As the computing industry hit the Power Wall over the last decade, applications need to be aware of their power consumption. In particular, they should use the hardware resources in a way that reduces power consumption without sacrificing performance. To achieve this balanced performance and power, a combined effort from software and hardware is necessary.;This dissertation presents a framework to achieve balanced performance and power for scientific applications on modern micro-processors by carefully orchestrating compiler optimizations to dictate the efficient utilization of hardware resources using these five main steps: 1) understand performance of scientific applications through performance models, 2) understand power consumption and thermal behavior of scientific applications through power and thermal models, 3) understand the impact of optimizations on performance, power, and temperature, 4) understand various application behaviors and categorize them, and 5) use machine learning techniques to automatically select optimizations for balanced performance and power. |
