Global Round Robin: Efficient Routing With Cut-Through Switching in Fat-Tree Data Center Networks

Fat tree is a scalable and widely deployed data center network topology. In this paper, a novel framework for designing per-packet load-balanced routing algorithms in fat tree called global round robin (GRR) is proposed. Routing in fat tree consists of uprouting and downrouting. In uprouting, a packet is sent to a switch that is a common ancestor (CA) of the source (server) and the destination. In downrouting, the packet is sent from the CA switch to the destination. Assume that time is slotted and each slot can accommodate one packet. With GRR, in each slot, a connection configuration is formed by establishing an uprouting path from each server to a spine switch port such that no paths will cross each other. Packets are sent from sources to respective spine switches with cut-through switching. The connection configuration is updated in a round robin fashion such that in every m slots, where $m$ is the number of spine switches, each server is connected to each spine switch exactly once. Since a CA does not need to be a spine switch, an improved GRR (IGRR) is then proposed to allow the nearest CA to intercept packets for downrouting. We prove that both GRR and IGRR can guarantee 100% throughput under a wide class of traffic. An analytical model is also constructed for studying their delay performance under uniform traffic. Finally, simulation results show that IGRR provides the best delay-throughput performance among all the existing per-packet load-balanced routing algorithms.