As increasingly parallel applications run on progressively parallel computer systems, our means and measures for evaluating application performance prove increasingly inadequate.
Systemwide metrics like instructions per cycle quantify average performance but reveal little about an applications real-time quality of service (QoS). For parallel programs in particular, these averaging metrics fail to capture performance fluctuations that occur because of competitive resource sharing between threads and frequent dependencies between them. We propose, as an alternative, to develop benchmarking tools that will measure QoS using application-specific metrics. Additionally, we hope to explore ways of using QoS metrics to optimize system resource utilization.

My interest in computer architecture was spurred by 6.004 and further bolstered by systems classes like 6.172. As an intern at Microsofts Windows division last summer, my role encompassed a performance focus akin to this project. My choice of topic for presenting a computer science concept to high school students in 6.UAT last semester also revolved around caching and optimizing the memory hierarchy.