Scalable and Load-balanced Data Center Multicast

Data center applications use multicast as an effective method to reduce bandwidth cost. However, traditional multicast protocols designed for IP networks are usually bottlenecked by the limited state capacity on switches. In this paper, we propose a scalable multicast solution on fat tree networks based on the observation that data center multicast traffic has strong heterogeneity. We propose to remove the multicast management logic from switches and use the SDN controller to manage multicast groups. The proposed Dual-structure Multicast (DuSM) determines elephant and mice groups according to their traffic amounts and treats them separately. For each elephant group, the controller installs multicast state to maintain multiple shared trees and the group traffic will be balanced evenly among the trees to avoid congestion. For mice groups, the controller applies state-free multicast that trades bandwidth capacity for state capacity, such as multicast-to-unicast translation. Our experiments using real multicast traffic data show that, DuSM can increase the multicast state capacity to support more number of groups by > 200% compared to IP multicast. DuSM also achieves better traffic balance among links than IP multicast.