Handling Elephant Flows in a Multi-tenant Data Center Network

Abstract

A virtualized data center is highly multifarious environment, shared among hundreds of co-located tenants hosting heterogeneous applications. High degree of virtual machine consolidation leads to diverse tra c dynamics with uneven tra c demands. Tenants' virtual machines generate a subset of elephants or mice ows that traverse the underlay fabric in aggregate, i.e., encapsulated in tunneling protocols such as VXLAN, NVGRE and STT. Elephant ows are long-lived and bandwidth intensive whereas mice ows are short-lived, latency-sensitive and highly bursty in nature. Both types of ows require di erent treatment from underlay fabric but encapsulation obfuscates the overlay tra c characteristics and demands. Existing approaches employed in data centers such as ECMP are either agnostic to elephant and mice

ows or have no visibility into virtual tra c which may be used to precisely detect, isolate and treat elephant ows di erently than mice ows. If elephant ows are not identi ed and addressed in aggregated virtual tra c, they may a ect mice ows generated from co-located applications, hence degrading application performance of co-located tenants. Therefore, there is a need to manage virtual tra c along with physical tra c in data centers. This work identi es and treats elephant ows belonging to di erent tenants on virtual network components such as virtual switches residing in the hypervisor. Virtual switch is at the optimal location to detect and handle elephant ows before they saturate the underlay network fabric and may be used to ameliorate the network performance with minimal overhead both on host and in-network. The proposed scheme has been tested against state of the art approaches such as Hedera and ECMP and has shown improved bisection bandwidth on a 4-array fat-tree interconnect.