Accurate Traffic Load Properties Verification in Data Center Networks

Abstract

This paper presents our experiences in deploying YU, Alibaba’s in-house traffic load verification system for wide area networks (WAN), in data center networks (DCN). We encounter two key challenges: the opaque ISP challenge, where the opaque nature of the configurations of peering ISPs results in low accuracy of inter-DCN route simulation, and the DCN hash configuration challenge, where ad hoc hash configurations of switches result in traffic polarization and, hence, low accuracy of symbolic traffic execution. We introduce two novel designs in QI, the next generation of YU, to address these challenges. First, we abstract peering ISPs as a virtual BGP router in our global inter-DC network and synthesize its configuration to be equivalent to that of ISPs, substantially simplifying the ingress route simulation for DCN. Second, we synthesize viable hash configurations for switches using constraint programming, simultaneously enabling polarization-free actual traffic forwarding and accurate symbolic traffic execution for DCN. QI has been deployed in the daily operation of Alibaba’s DCN and has successfully identified multiple traffic load violation scenarios in DCN. We share representative cases with the community and evaluation results of QI from our production network, as well as our learned lessons and visions.