随机需求下面向异质费用的云资源调度算法

采用分布式云构建流媒体服务等高资源消耗系统,既符合应用多区域部署的要求,也能充分利用云中资源保证服务质量,同时还能进行系统预算成本控制.由于各区域云中心费用函数存在差别,分布式云中调度需引入异质费用模型,结合流媒体应用中用户请求高度动态随机的特征,在给定的费用预算下响应尽可能多的用户请求.均值需求模型忽略了资源需求在短时间间隔内的变化细节,导致资源利用率低下.为克服均值需求模型的缺点,采用随机需求模型以捕捉细粒度资源需求,使用通用代价函数描述异质费用模型,建立更具通用性的非线性规划问题模型;为降低求解算法的复杂度,基于动态规划快速获得解的下界,再迭代逼近获取近优解.实验结果表明:相比经典的基于均值的调度算法,在区域数量较大时,平均能额外满足15%的用户请求;随着预算的减少,能额外满足近40%的用户请求;且不受各区域价格函数差异和用户访问需求差异的影响.因此,在构建全球部署的大规模流媒体服务系统时,算法能以较低的计算代价显著增加响应的用户请求量,广泛适应各种不同的云基础设施服务提供商.

Heterogeneous cost oriented cloud resource scheduling algorithm for stochastic demand

It is recommended to use a distributed cloud to construct a high resource consumption system such as streaming media service, which not only meets the requirements of multi-region deployment, but can also make full use of the resources in the cloud to ensure the service quality while controlling system budget costs. Make the difference of cost functions in cloud centers sitting in different regions, the heterogeneous cost model should be introduced in distributed cloud oriented scheduling. Given the highly dynamic and random characteristics of user requests in streaming media applications, it is expected to respond to as many user requests as possible under a given cost budget. Mean demand model ignores details of changes in resource requirements over short time intervals, and leads to inefficient use of resources. To overcome the disadvantages of mean demand model, we use a stochastic model to capture the fine-grained information of resource demand, and use a common cost function to describe the heterogeneous cost model, and also establish a general nonlinear programming problem model. To reduce the algorithm complexity, the lower bound of the solution is quickly obtained based on dynamic programming, and then the near optimal solution is obtained by iteratively approximation. The simulation results show that, compared with the classical average demand model based scheduling algorithm, the proposed algorithm can additionally meet 15% more of the user requests when the number of regions is large, and can additionally satisfy up to nearly 40% of the user requests as the budget decreases. The algorithm is not affected by differences in price functions or by user requests statistics across different regions. As a result, when used in globally deployed large-scale streaming media system, the proposed algorithm can significantly increase the number of satisfied user requests with limited computing time, thus is adaptable to a wide range of cloud infrastructure service providers.