variability in the execution of multimedia applications and implications for general-purpose architectures

Multimedia applications are an increasingly important workload for a large range of systems in-cluding handheld computers, laptops, and desktops. It is expected that general-purpose processors will be increasingly employed for such workloads; however, there are still some shortcomings that they need to overcome. One shortcoming of general-purpose architectures results from their use of complex techniques (e.g., out-of-order issue, dynamic branch prediction, and caches) for high performance. It is conjectured that these complex features result in highly unpredictable execution times, making these architectures unsuitable for meeting the real-time requirements of multimedia applications. There is, however, little quantitative data to support or disprove this conjecture. This thesis performs a detailed quantitative analysis of execution time variability at the frame granularity for a number of multimedia applications processing speech, video, and audio data types. We nd that while execution time varies from frame to frame for many applications, this variability is mostly caused by the algorithms used by the applications and the media input itself. Aggressive architectural features induce little additional variability (and unpredictability) in execution time, in contrast to conventional wisdom. Our nding of frame-level execution time variability also motivate