Flow-Level Rerouting in RDMA-Enabled Dragonfly Networks

Due to the characteristic of large-radix routers, the Dragonfly topology can achieve low diameter, high performance/cost ratio. However, in the Dragonfly networks deployed with Remote Direct Memory Access (RDMA), existing packet-level routing algorithms which are mostly based on queue length information, are neither good enough to achieve load balancing nor meet the requirement of in order. To tackle the above issues, we first analyze the drawbacks of flow-level source routing in RDMA-enabled Dragonfly networks. Then, a flow-level rerouting scheme that can estimate traffic distribution and link load based on the routers' history information is proposed. Finally, the simulation results show that our scheme can obtain significant performance gains over existing algorithms in both average flow completion time (AFCT) and saturation throughput. In particular, under the adversarial traffic pattern, our scheme can greatly reduce the AFCT of flow-level UGAL by 25% and improve the saturation throughput by 13% while avoiding disorder.