CAJ | 학술논문

　　MapReduce是云计算中重要的批数据处理框架，多任务共享MapReduce机群并满足任务实时性要求是调度算法急需解决的问题。提出两阶段实时调度算法，将调度划分为任务间调度和任务内调度。对于任务间调度，使用抽样法和经验值法确定子任务执行时间，利用该参数建立资源分配模型，动态确定任务优先级进行调度；对于子任务使用延迟调度策略进行调度，保证计算的本地性。实验结果显示，两阶段实时调度算法相比公平调度算法和FIFO算法，在保证吞吐量的同时能够满足任务实时性要求。
　　MapReduce시운계산중중요적비수거처리광가，다임무공향MapReduce궤군병만족임무실시성요구시조도산법급수해결적문제。제출량계단실시조도산법，장조도화분위임무간조도화임무내조도。대우임무간조도，사용추양법화경험치법학정자임무집행시간，이용해삼수건립자원분배모형，동태학정임무우선급진행조도；대우자임무사용연지조도책략진행조도，보증계산적본지성。실험결과현시，량계단실시조도산법상비공평조도산법화FIFO산법，재보증탄토량적동시능구만족임무실시성요구。
MapReduce is a popular batch data processing framework in cloud computing field. Sharing MapReduce cluster and meeting the deadlines of jobs is a key problem to be solved. This paper proposes a two phase real-time scheduling algorithm which separate scheduling into job scheduling and task scheduling.It uses sampling method to estimate the task excuting time so that the scheduler can make a decision on how many slots should be assigned to the job and how to calculate the job's priority. Using delay-scheduling scheme in task scheduling, the"computing locality"problem can be solved well. Experiments result shows that the scheduling algorithm implemented in this paper satisfies the job's real-time requirement as well as throughput of the cluster.