Modeling multimedia workloads for embedded system design

To design a successful computer system, designers need to kn ow characteristics of the computational workload that this system is supposed t o process. This knowledge forms the necessary basis for optimizations of th e system. In order to use this knowledge in the design process, designers need t o characterize the workload using a formal workload model. This model represen t a abstraction of the concrete workload and serves as an input to a number of c ritical design tasks, such as system performance analysis. The quality of t he workload model largely determines the quality of the design decisions made b s d on it. Coming up with a proper workload model represents a difficult p roblem in many computer system design contexts. One such context, add ressed in this thesis, is system-level design of embedded computers whose mai n functionality involves real-time processing of media streams (e.g. streams of audio-video data). Of late, there is a growing demand for such computers because they are increasingly being embedded into many electronic products, especi ally those found in consumer electronics domain, e.g., digital TVs, audio and v ideo players, digital video cameras, advanced set-top boxes, multimedia-enable d mobile phones and a myriad of other electronic devices supporting multimedia applications. To meet high performance requirements and stringent constraints p er aining to cost, size and energy consumption, these embedded computers tend to ha ve complex, heterogeneous multiprocessor architectures. This architect ural complexity, coupled with the ever-growing complexity of the multimedia applica t ons themselves, results in a very complex workload behavior and by that poses ma ny challenges to the workload modeling. In this thesis, we argue that the variability of various para meters of the multimedia workloads is the key property to be captured in a worklo ad model for the embedded systems design. We show that conventional workloa d models fail to accurately characterize the dynamic nature of the multimed ia workloads and, as a result, return overly pessimistic estimations of system p rformance (especially, if worst-case performance bounds are of interest). As a solu tion, we propose a novel workload model capable of accurately capturing the wo rkload’s dynamic nature. We demonstrate the advantages of the proposed workl oad model over conventional ways to characterize the workload and develop a number of systemlevel design methods which use this model. These methods inc lude system-level performance analysis, automatic identification of represe ntative workload sce-