| 摘要: |
| 目前大数据处理过程较少关注任务所处理数据间的依赖关系，在任务执行过程中可能产生大量数据迁移，影响数据处理效率.为减少数据迁移，提升任务执行性能，从数据关联性及数据本地性两个角度出发，提出了一种数据关联性驱动的大数据处理任务优化调度方案：D3S2（data-dependency-driven scheduling scheme）.D3S2由两部分组成：（1）数据关联性感知的数据优化放置机制（dependency-aware placement mechanism，简称DAPM），根据日志信息挖掘数据关联性，进而将强关联的数据聚合并放置于相同机架上，减少了跨机架的数据迁移；（2）数据迁移代价感知的任务优化调度机制（transfer-aware scheduling mechanism，简称TASM），完成数据放置后，以数据本地性为约束，对任务进行统一调度，最小化任务执行过程中的数据迁移代价.DAPM和TASM互相提供决策依据，以任务执行代价最小化为目标不断迭代调整调度方案，直至最优任务调度方案.在Hadoop平台上进行的实验结果表明：较之原生Hadoop，在不增加作业完成时间的基础上，D3S2减少了作业执行过程中的数据迁移量. |
| 关键词: 数据关联性 数据本地性 数据放置 任务调度 迁移代价感知 |
| DOI：10.13328/j.cnki.jos.005236 |
| 分类号: |
| 基金项目:国家高技术研究发展计划（863）（2013AA01A209）；国家自然科学基金（61172048，61303250） |
|
| Dependency-Driven Task Scheduling Scheme of Big Data Processing |
|
WANG Bin, WU Ya-Jing, YANG Xiao-Long, SUN Qi-Fu
|
|
School of Computer and Communication Engineering, University of Science and Technology Beijing, Beijing 100083, China
|
| Abstract: |
| Currently, there is lack of consideration of dependencies between data in big data processing, resulting in low data processing efficiency with large amounts of data transfer during task execution. In order to reduce data transfer and improve processing performance, this paper proposes a data-dependency driven task scheduling scheme, named D3S2, for big data processing. D3S2 is mainly composed of two parts:dependency-aware placement mechanism(DAPM), and transfer-aware task scheduling mechanism(TASM). DAPM discovers dependency between data so that strongly related data will be clustered and assigned to nodes in the same rack, thereby reducing the cross-rack data migration. TASM schedules tasks simultaneously after data placement according to the data locality constraint, so as to minimize the data transfer cost during the task execution. DAPM and TASM provide basis for decision making to each other, iterating constantly to adjust the scheduling scheme with the goal of minimizing the execution cost until an optimal solution is reached. The proposed scheme is verified in Hadoop environment. Experiments show that compared to native Hadoop, D3S2 reduces the data transfer during job execution, and shortens job running time. |
| Key words: data-denpendency data-locality data placement task scheduling transfer-cost-aware |
