Efficient Traffic Splitting on SDN Switches

Networks often need to balance load over multiple backend servers—or other components like links, paths, or middleboxes—offering the same service. For example, a large cloud provider could easily host tens of thousands of services, each with dozens or hundreds of backend servers. Hash-based approaches like Equal-Cost Multi-Path (ECMP) can achieve a roughly uniform split, or even a weighted split by repeating the same “next-hop” multiple times. However, ECMP has low accuracy for approximating arbitrary weights, scales poorly with many “next-hops,” and experiences significant churn when weights change. Instead, a Software-Defined Network (SDN) could split traffic more flexibly by crafting a set of wildcard rules to install in OpenFlow switches. However, existing SDN-based loadbalancers either send data packets to the controller, or generate too many rules to approximate the weights. In this paper, we propose Niagara, an SDN-based load-balancing scheme where the controller computes an efficient approximation of the weights for each service, and optimizes the division of the rule-table space across multiple services. Our experiments demonstrate that Niagara improves traffic-splitting accuracy by 77% compared to ECMP, while requiring only 19% of the rules used by Weighted-Cost Multi-Path. Niagara requires 63% fewer rules than previous SDN approaches.