云环境下面向跨域作业的调度方法

A Scheduling Strategy for Jobs Across Geo-Distributed Datacenters in Cloud Computing

云环境下,因数据局部性或是任务对资源的特殊偏好,一个作业所包含的任务往往需要在不同的数据中心局点上运行,此类作业称为跨域作业.跨域作业的完成时间取决于最慢任务的执行效率,即存在木桶效应.针对各域资源能力异构条件下不合理的调度策略导致跨域作业执行时间跨度过长的问题,本文提出一种面向跨域作业的启发式调度方法 MIN-Max-Min,优先选择期望完成时间最短的作业执行.通过实验表明,与先来先服务的策略相比,该方法能将跨域作业平均执行时间跨度减少40%以上.

In cloud computing,tasks in a job often need to run on different datacenters due to the input data locality or special preference for resources,that is,the job runs across geo-distributed sites. The different tasks in a job have to be scheduled in different domain（ data center） to execute for their personalization requirements,so the job completion time depends on the slowest task,which is called＂barrel effect＂. As geo-distributed scheduling strategy without regard to heterogeneous resources leads too long execution time span,this dissertation proposes an optimization strategy for geo-distributed scheduling named M IN-M ax-M in. The strategy gives priority to select the expectation shortest completion job to execute by heuristic rule. Experiments showthat compared with first come first service strategy,the strategy can reduce cross domain average execution time span to less than 40% under the simulation load.