Adaptive flow scheduling for modular datacenter networks

Abstract

Modular data center networks leverage recursive topology to speed up several traffic patterns and increase the network capacity for various data shuffle applications. However, the applications of online social networks and instant messaging may be deployed in the same data center for high cost-effective investment. The large amount of mice flows which produced from these applications may cause partial path congestion, and these congestion in turn can prevent the mice flow access that reduces message loss or delay for online social networks. In this stydy, we define the cost function for different parallel paths and model the path selection process, and propose an adaptive flow scheduling for modular data center networks (AFMD) to search more suitable path for both large flows and mice flows with the greedy algorithm. The experiment shows that AFMD can improve large flow throughput by up to 25.9 % and the mice flow access ratio by up to 16.5 % and AFMD exhibits effective complementarity and compatibility to BCube networks.