数据操作系统中的双层优化调度模型

调度问题是数据操作系统研究中的关键性问题,它建立了计算资源、计算任务以及数据间的链接关系。在海云协同网络环境下的调度问题中,常见的调度考量包括公平性、数据本地性等。由于数据操作系统使用环境的演化,工作负载中任务的交互特性给调度问题提出了新的挑战。本文在保留传统调度考量的基础上,兼顾交互、批处理两种作业模式的异同,提出一种优化的双层调度模型,并使用符合实际产业环境分布的工作负载在现实集群上对该调度模型进行了验证。实验结果说明,该模型以微量降低系统吞吐量的代价整体优化了交互作业的响应时间,同时兼顾了用户级公平性。

An Optimized Two-level Scheduling Strategy in DataOS

In DataOS,scheduling is the key mechanism that connects computing resources,jobs and data. Scheduling concerns in the current sea-cloud coordinating networked environment includes fairness concern,locality concern,etc. More importantly,due to the evolvement of workload in DataOS,the interactive behavior in workload post a new challenge for system scheduler. In this paper,based on traditional scheduling concerns and the new workload structure,a new two-level optimal scheduling model is presented. Results shows that with the new scheduling model,response time of interactive jobs are improved with the cost of small reduce in system throughput.