On traffic fairness in data center fabrics

A present challenge in data center networks (DCN) is to better understand the impact of novel flattened and modular DCN architectures on congestion control protocols, and viceversa. One of the major concerns in congestion control being the fairness in the offered throughput, the impact of the additional path diversity and forwarding features, brought by the novel DCN architectures and protocols, on the throughput of individual endpoints (servers) and aggregation points (edge switches) is unclear. This paper attempts to answer these open questions. Specifically, how best is the allocation of the competing elastic demand flows and how is this allocation impacted with the increase in capacity? We provide an optimization formulation of the problem based on the proportional fairness principle of TCP. We conducted a series of test scenarios on the fat-tree data center topology by considering load balancing and link capacities for different network cases in order to present our analysis on the results. We observed that the traffic allocation fairness is primarily impacted by the weights associated with the TCP implementation in use.