VOD系统批处理调度策略优化研究

Optimization on Batching Schedule Policy for Video-On-Demand System

大规模视频点播（VOD）业务需要消耗大量的网络带宽和服务器资源.基于多点组播的批处理技术允许服务器在不增加额外系统资源的情况下发送相同数据到多个客户端,从而达到系统资源利用率的最大化.通常使用的批处理调度策略有FCFS、MQL和MFQL三种,但是这些方案没有得到令人满意的结果.本文提出一种新策略,称之为MDP,即最大撤销概率,其基本思想是在组空间搜索用户请求撤销概率和的最大值.MDP以等待服务的所有组中撤销概率的和最大为调度决策依据,保障到达系统的用户因不能忍受等待时间太长而退出的概率最小.通过分析与仿真实验,MDP方案在系统吞吐量、平均等待时间、用户请求撤销概率和不公平性等系统性能参量方面获得优良结果.

The paper proposes a novel scheduling scheme, called Maximum Defection Probability i. e. MDP, which makes a decision by the maximum of the whole latency or the maximum of defection probability among all the groups waiting for service. The basic idea of MDP is to search the maximum of the sum of user requests defection probability among all the groups. In the VOD system, MDP is able to schedule a idle channel to firstly serve for the batch group with maximum defection probability. Therefore the MDP can minimize the probability that the users want to renege their on-demand for a video because of overfull latency. This policy adaptively adjusts itself to cope with a changing workload including queue length and their latency. A simulation model with Zipf distribution is developed to compare the proposed MDP variants with FCFS, MQL and MFQL. The simulations and authors' analysis show that MDP yields excellent empirical results in terms of standard performance measures such as throughput, user requests defection probability, average latency and unfairness. The batching policy of MDP is easy to implement in practical VOD systems without a priori knowledge of any traffic pattern.