Alleviating network congestion for HPC clusters with fat-tree interconnection leveraging software-defined networking
Zhenwei Wu, Kai Lü, Xiaoping Wang, Wanqing Chi · 2016
Interconnection networks are essential elements in nowadays High Performance Computing (HPC) clusters, which must meet the urgent requirement for fast and reliable communication from the HPC applications. As a consequence, the network performance have a great influence on the overall system one. However, traditional routing mechanisms and load-balancing approaches cannot make efficient use of the network due to lack of network status information and flexible ways to perform dynamic controlling operations. As a solution, the emerging Software-Defined Networking (SDN) network paradigm, offers great opportunities to enhance the network performance with its low-cost traffic statistics monitoring mechanism and centralized network controller. In this paper, we present a congestion-aware SDN controller for Fat-tree interconnection network. We implement a network congestion detector based on the traffic monitoring mechanism in SDN. In addition, we design a traffic flow scheduler which is capable of dynamically redistributing the network traffics flowing along the congested path to the congestion-free ones. The Ryu SDN Framework is employed in our implementation. Evaluation experiment conducted on the Mininet network emulator shows that our controller could efficiently alleviate the network congestion and achieve better use of the target network.