VirtMonE: Efficient detection of elephant flows in virtualized data centers

A modern virtualized data center is highly multifarious environment shared among hundreds of co-located tenants hosting heterogeneous applications. The tenants´ virtual machines generate a subset of elephants or mice flows (different in terms of rate, size, duration, and burstiness) based on the type of application they are running. Virtual traffic generated from the tenant´s virtual machines traverses the underlay physical fabric in aggregate because of different encapsulation techniques (VXLAN, NVGRE, and STT for example) employed in data center networks thus obfuscating the virtual traffic characteristics. Existing approaches to monitor and/or identify elephant flows either have limited or no visibility into virtual traffic or are associated with high monitoring overhead making it hard to precisely detect and properly engineer elephant flows on the underlay fabric. In this paper, we present VirtMonE, a lightweight detection mechanism aimed at precisely detecting egress elephant flows at Open vSwitch while providing visibility into virtual traffic with least measurement and monitoring overhead at the edge. We conduct simulations on a small GENI testbed to evaluate the performance of the proposed solution for a software-defined multi-tenant virtual network. Our proposed solution is demonstrated to precisely detect the elephant flows from different tenants at the edge, provide visibility into virtual traffic and mitigate the network overhead associated with detection, thus improving the overall performance of the data centre.