Joint scheduling of tasks and network flows in big data clusters

Abstract

As an increasing number of big data processing platforms like Hadoop, Spark, and Storm appear and normally share the resources in the data center, it has been important and challenging to schedule various jobs from these platforms onto the underlying data center resources such that the overall job completion time is minimized. To solve the problem, the existing work either focus on the task-level scheduling techniques, such as Quincy and delay scheduling, or focus on the network flow scheduling techniques, such as D3 and preemptive distributed quick. These works deal with the scheduling of tasks and network flows separately and cannot achieve optimal performance. The reason is that the task scheduling without regard of the available network bandwidths may generate the task placement that causes serious network congestions and thus leads to long data transmission time. In this paper, we propose the joint scheduling technique by coordinating the task placement and the scheduling of network flows arising from these tasks. We develop a software-defined network (SDN)-based online scheduling framework which selects the task placement based on the available bandwidth on the SDN switches and at meanwhile optimally allocates the bandwidth to each data flow. Comprehensive trace-driven simulations show that the joint scheduling technique can take full use of the network bandwidth and thus reduce the job completion time by 55% on average compared with the benchmark methods.