低消費電力マルチコアプロセッサとその並列ソフトウェアに関する研究

点击翻译按钮获取中文摘要

The research lead to new results in the area of power control for mobile devices and for (cloud) servers which are found in data centers. On the Linux- and Android-operating system the run-time of OSCAR-compiled applications has been improved and also became more predictable. In order to achieve this success the operating system was modified to support more deterministic task to processor mappings. Furthermore the operating system was extended with a new programming interface (API) to control the performance of the processor cores. The API includes operations for dynamic frequency and voltage scaling (DVFS), clock- and power-gating. To assess the performance the operating system also collects data on CPU- and DRAM power consumption as well as processor temperature. The measurement setup performs fully automatic data acquisition. This allowed thousands of measurements to be collected over a month. The data was used to perform an offline design space exploration for a multi-core compiler run-time ("libgomp") and a scientific application. As a result the power consumption could be reduced without any modifications to the binary code of the application.