CSR--EHS: Dynamic Resource Management for Multimedia Applications on Embedded Systems

在嵌入式系统上实现多媒体编解码算法的现有方法具有几个缺陷。首先，多媒体算法的设计没有显式地考虑资源约束。其次，当前的嵌入式系统在设计和实现压缩技术时往往假设资源利用率是最差的，从而忽略了多媒体算法需要时变资源的事实。第三，算法及其实现没有联合优化，导致性能次优。最后，现有的多媒体自适应系统充其量是被动的，即它们使其资源适应过去的数据输入，而不是相应地建模和预测未来的资源需求和规划。在本项目中，为了提高嵌入式系统上多媒体应用的效率，本项目针对上述局限性进行了研究。具体地说，本研究探讨动态资源规划技术，该技术可以在满足给定系统约束的情况下，联合优化多媒体算法的复杂度和资源利用率。通过使用随机建模、学习和预测等前馈控制机制，这形成了主动资源管理的基础。最后，速率和复杂度可扩展的压缩方案对于实现对可用资源的优化适配是必要的。开发这样一个集成的应用程序实现框架具有根本的重要性，因为它不仅可以提高与现有嵌入式系统相比的多媒体性能，还可以为下一代多媒体压缩算法的设计以及应该具有多媒体意识的嵌入式系统设计提供有价值的见解。更广泛的影响体现在教育下一代学生和行业合作伙伴资源感知系统设计方面，这对实现未来泛在和自主计算系统的愿景至关重要。

Existing approaches to implementing multimedia encoding and decoding algorithms on embedded systems have several drawbacks. First, multimedia algorithms are designed without explicitly taking resource constraints into account. Second, current embedded systems often assume "worst-case" resource utilization for the design and implementation of compression techniques, thereby neglecting the fact that multimedia algorithms require time-varying resources. Third, the algorithms and their implementation are not jointly optimized, resulting in sub-optimal performance. Finally, existing adaptive systems for multimedia are, at best, reactive, i.e., they adapt their resources to past data inputs, instead of modeling and predicting future resource requirements and planning accordingly. In this project, the abovementioned limitations are addressed to improve the efficiency of multimedia applications on embedded systems. Specifically, this research investigates dynamic resource planning techniques that can jointly optimize the complexity of the multimedia algorithms and the resource utilization, while meeting a given system constraint. This forms the basis for proactive resource management by using feed-forward control mechanisms like stochastic modeling, learning and forecasting. Finally, rate- and complexity-scalable compression schemes are necessary to enable optimized adaptation to available resources. Developing such an integrated application-implementation framework is of fundamental importance, since it not only leads to improved multimedia performance over existing embedded systems, but also provides valuable insights into the design of next generation multimedia compression algorithms that should be complexity scalable, as well as embedded systems designs that should be multimedia aware.The broader impact is in the area of educating next generation students and industry partners in resource-aware system design, which is critical to realize the vision of ubiquitous and autonomous computing systems of the future.