On capacity provisioning in datacenter networks for full bandwidth communication

Recent advances in datacenter network design have enabled full bandwidth communication based on the notion of Valiant load balancing. In this paper, we study from the link capacity viewpoint how such communication paradigm can be supported in the context of link failures, the study of which is currently absent. In particular, we target full bandwidth communication among all the servers, for all valid traffic patterns, and under k arbitrary link failures. We derive the minimum link capacity required on two typical datacenter topologies-VL2 and fat-tree. Our main finding is that given the same server scale, fat-tree requires less total link capacity than VL2 for 1 ≤ k ≤ n/4, where n denotes the port count of homogeneous switches used in fat-tree. For k > n/4, there exists a turning point from which VL2 becomes more capacity-efficient.