|Luping Wang||Hong Kong University of Science and Technology, Hong Kong|
|Wei Wang||Hong Kong University of Science and Technology, Hong Kong|
|Bo Li||Hong Kong University of Science and Technology, Hong Kong|
Performance and service isolation come as two top objectives for coflow scheduling. However, the common wisdom is that these two objectives are often conflicting with each other and cannot be achieved simultaneously. Existing coflow scheduling frameworks either focus only on minimizing the average coflow completion time (CCT) (e.g., Varys), or providing optimal isolation between contending coflows by means of fair network sharing (e.g., HUG). In this paper, we make an attempt to achieve the best of both worlds through a novel coflow scheduler, Utopia, to attain near-optimal performance with provable isolation guarantee. This is particularly challenging given the correlation of bandwidth demands across multiple links from coflows. We show that Utopia is capable of reducing the average CCT dramatically, while still guaranteeing that no coflow will ever be delayed beyond a constant time than its CCT in a fair scheme. Both trace-driven simulation and EC2 deployment confirm that Utopia outperforms the fair sharing policy by 1.8× in terms of average CCT, while producing no completion time delay for a single coflow. Even compared with performance-optimal Varys, Utopia speeds up average coflow completion by 9%.