An experimental investigation of Round-Trip Time and virtualization

Virtualization is omnipresent in our networks and data centers. Little is known of its overhead, especially regarding the deployment of Virtual Network Functions and the use of existing network active measurement tools. This paper evaluates the impact on Round-Trip Time of parameters such as CPU affinity, the frequency of injection measurements echo packets, the type of virtual network driver, the use of CPU, I/O or network overload, and the number of background VMs. We compare RTT results using three virtualization technologies, namely, KVM, LXC, and Docker containers. To isolate the role of virtualization on packet timing overhead, we also include kernel-level measurement points in our experiments. We discover that the parameters differ in their impact on RTT, and that their effect changes according to how they are combined. The results reported in this paper allow a Cloud administrator to understand the overhead of virtualization technologies such as KVM, LXC, and Docker containers as well as the impact of virtual network drivers, disk I/O overhead, CPU processing and combinations therein, on network performance.