Efficient techniques for detecting and exploiting runtime phases

Most applications have time-varying runtime phase behavior. For example, alternating between memory-bound and compute-bound phases. Nonetheless, the predominant approach in computer research has been to categorize an application based on its average runtime behavior. However, this can be misleading since the application may appear to be neither memory nor compute bound. In this thesis we introduce tools and techniques to enable researchers and software developers to capture the true time-varying behavior of their applications. To do so, we 1) develop an efficient technique to detect runtime phases, 2) give new insight into applications' runtime phase behaviors using this technique, and, finally, 3) explore different ways to exploit runtime phase detection. The results are ScarPhase, a low-overhead phase detection library, and three new methods for exploring applications' phase behaviors with respect to: 1) cache performance as a function of cache allocation, 2) performance when sharing cache with co-running applications, and finally, 3) performance and power as a function of processor frequency. These techniques enable us to better understand applications' performance and how to adapt different settings to runtime changes. Ultimately, this insight allow us to create new faster and more power efficient applications and runtime systems that can better handle the increasing computation demands and power constraints of tomorrow's problems.UPMAR