Towards a systematic approach for modeling and optimizing distributed and dynamic multimedia systems

Recent advances in low-power, multi-core and distributed computing technologies have opened up exciting research opportunities, as well as unique challenges, for modeling, designing, and optimizing multimedia systems and applications. First, multimedia applications are highly dynamic, with source characteristics and workloads that can change significantly within milliseconds. Hence, systems need to be able to optimally adapt their scheduling, resource allocation, and resource adaptation strategies on-the-fly to meet the multimedia applications' time-varying resource demands within the delay constraints specified by each application. Second, systems often need to support multiple concurrent multimedia applications and thus, (Pareto) efficient and fair resource management solutions for dividing processing resources among the competing applications need to be designed. Finally, some applications require distributed computing resources or processing elements, which are located across different autonomous sites. These different sites can collaborate in order to jointly process the multimedia data by exchanging information about their specific system implementations, algorithms and processing capabilities. However, exchanging this information among these autonomous entities may result in unacceptable delays or transmission overheads. Moreover, they may even refuse to share this information due to proprietary or legal restrictions. Thus, information-decentralization can present a major obstacle for optimizing the performance of delay-sensitive multimedia applications that require coordination and cooperation between distributed, autonomous sites.;This dissertation addresses the above challenges by providing a systematic framework for modeling and optimizing multimedia systems in dynamic, resource-constrained, and informationally-distributed environments. In particular, we propose a stochastic modeling approach to capture the dynamically changing utilities and workload variations inherent in multimedia applications. This approach enables us to determine analytical solutions for optimizing the performance of applications on resource-constrained systems. Furthermore, the problem of information-decentralization can be addressed in our framework by systematically decomposing the joint multi-applications and multi-site optimization problems, and designing corresponding mechanisms for exchanging model parameters, which characterize the utilities, constraints and features of the autonomous entities. This systematic decomposition enables entities to autonomously coordinate and collaborate under informational and delay constraints. Finally, to optimize the performance of the multimedia applications or systems in these distributed environments, we deploy multi-agent learning strategies, which enable individual sites or applications to model the behaviors of its competitors or peers and, based on this, select their optimal parameters, configurations, and algorithms in an autonomous manner. Summarizing, our framework proposes a unified approach combining stochastic modeling, systematic information exchange mechanisms, and interactive learning solutions for optimizing the performance of a wide range of multimedia systems.;A unique and distinguishing feature of our approach is the extent of multimedia algorithms and systems domain specific knowledge used in developing the proposed framework for modeling, and optimizing the interacting system components and applications. This is in contrast to existing distributed optimization or game theoretic approaches, which use simplistic utility---resource functions, and often ignore the dynamics and constraints experienced in actual multimedia systems. Instead, our developed modeling and optimization framework is directly shaped by the specific characteristics, constraints and requirements of multimedia systems. Specifically, the proposed framework provides pragmatic implementation solutions for (i) the optimization of dynamic voltage scaling algorithms for multimedia applications, (ii) energy-aware resource management for multiple multimedia tasks, and (iii) resource-constrained adaptation for cascaded classifier topologies in distributed stream mining systems.