System software performance monitoring and optimization

点击翻译按钮获取中文摘要

Our primary objective is to make improvements to operating systems so as to significantly improve kernel and application performance. Currently, we intend to target multicore-based systems, because the speed at which multicore systems became widespread has caught the software community by surprise, and as a result, most of these systems operate far below their potential. Our approach is to exploit Hardware Performance Counters (HPCs) that today are an integral part of all processors' Performance Monitoring Units (PMUs) and use them to measure overheads and identify bottleneck online and in real time. We then feed the information gathered from the HPCs to the system resource managers so that they can make more informed decisions on how to best use the system resources from a performance point of view. This results in a continuous feedback optimization loop. The work we do is experimental and implementation oriented. We design and implement systems, or modify existing systems, such as Linux. Real-world workloads are used in our experiments.