Network-aware task selection to reduce multi-application makespan in cloud

Abstract One new metric that plays a vital role in evaluating the cloud service is the multi-application makespan. There are usually multiple applications without a deadline in the cloud, while the makespan of each application should be minimized, such as parallel applications for training neural networks. However, the previous scheduling rule, called moving the computation tasks of applications closer to data, fails to reduce multi-application makespan. Because it is unknown which application the task belongs to, the number of optimized tasks (close-to-data tasks) is unbalanced among multiple applications, which causes a big gap between the makespan of multiple applications. To address this issue, we propose a cooperative scheduler that considers the application-task relationship, whose goal is reducing the multi-application makespan. Firstly, we propose a cost sharing game model to guide the balance of optimized task selection between multiple applications, in which the cost is referred to as the makespan. Specifically, we develop a network-aware method that can accurately estimate the makespan with a given task selection strategy. In the end, we evaluate the cooperative scheduler in the actual Hadoop cluster with diverse network environments. Experimental results demonstrate that the gap between multi-application makespan decreases. Moreover, compared to the baseline schedulers, the longest makespan decreases by 61.5%, and the network traffic is saved by more than 50%.