Energy-efficient scheduling of compute-intensive video processing tasks

视频服务的功耗目前是一个严重的问题，随着用户数量肯定会继续快速增长，这一问题只会增加(从2010年到2015年，移动数据流量预计将增长40倍，而移动视频预计将占美国流量的68.5%)。此研究计划的主要目标是调查并发现显著降低计算密集型视频处理任务(如视频压缩和转码)能耗的解决方案。申请者和他的团队将通过三项研究活动实现这一目标。前者将创建新的方法和软件工具来衡量视频任务的性能(计算复杂性、能耗和视频质量)如何受到视频处理参数(例如运动估计算法)、视频内容参数(例如帧数量、分辨率)和平台参数(例如处理器类型)的影响。第二部分将对结果进行分析，并建立数学模型，根据这些参数来预测视频任务的性能。本研究将利用各种机器学习技术(如聚类和支持向量机)来阐述这些视频处理性能模型。第三个将创建一个能效高的调度器，能够在计算机集群上近乎最佳地分配视频处理任务。最优标准将基于视频质量和功耗之间的理想折衷，受完成任务的时间、最低视觉质量和最大功耗等限制。视频处理性能模型对于提高调度器性能至关重要，因为要确定在满足约束的同时提供最佳折衷的处理参数。这个项目将利用最新的优化方法来创建这个节能的调度器。该团队将在两家视频公司的服务器上验证他们的创新，并与他们在其他研究项目上进行合作。这项重点关注电力消耗的新提议，如果成功，可能会产生一些想法，推动他们的合作伙伴进一步发展。

Power consumption for video services is currently a serious concern that will only increase as the number of users is certain to continue its rapid rise (mobile data traffic is expected to grow 40-fold from 2010 to 2015, and mobile video is expected to account for 68.5% of that traffic in the U.S.). The main objective of this research program is to investigate and discover solutions to significantly reduce the energy consumption of compute-intensive video processing tasks, such as video compression and transcoding. The applicant and his team will achieve this objective through three research activities. The first will create novel methodologies and software tools to measure how a video task's performance (computational complexity, energy consumption and video quality) is affected by video processing parameters (e.g. the motion estimation algorithm), video content parameters (e.g. number of frames, resolution) and platform parameters (e.g. type of processor). The second will analyze the results and elaborate mathematical models to predict a video task's performance as a function of these various parameters. Various machine learning techniques (such as clustering and support vector machines) will be exploited in this research to elaborate these video processing performance models. The third will create an energy-efficient scheduler capable of near-optimally distributing video processing tasks on a computer cluster. The optimality criterion will be based on the desired compromise between video quality and power consumption, subject to constraints such as the time to complete the task, minimum visual quality and maximum power consumption. The video processing performance models will be crucial for improving scheduler performance, in terms of identifying the processing parameters that provide the best compromise while at the same time meeting the constraints. The latest optimization methods will be exploited in this project to create this energy-efficient scheduler. The team will validate their innovations on servers from two video companies with which they collaborate on other research projects. This new proposal, which focuses on power consumption, could, if successful, generate ideas to propel their partners still further.