Performance Optimization of Hypervisor’s Network Bridge by Reducing Latency in Virtual Layers

Abstract

Multi tenant enabled cloud computing is the de facto standard to rent computing resources in data centers. The basic enabling technology for cloud computing is hardware virtualization (Hypervisor), software defined networking (SDN) and software defined storage (SDS). SDN and SDS provision virtual network and storage elements to attach to a virtual machine (VM) in a computer network. Every byte processed in the VM has to travel in the network, hence storage throughput is proportional to network throughput. There is a high demand to optimize the network throughput to improve storage and overall system throughput in big data environments. Provisioning VMs on top of a hypervisor is a better model for high resource utilization. We observed that, as more VMs share the same virtual resources, there is a negative impact on the compute, network, and storage throughput of the system because the CPU is busy in context switching (Popescu et al. https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-914.pdf [1]). We studied KVM (Hirt, KVM—The kernel-based virtual machine [2]) hypervisor’s network bridge and measured throughput of the system using benchmarks such as Iometer (Iometer, http://www.iometer.org/ [3]), (Netperf, http://www.netperf.org/netperf/NetperfPage.html [4]) against varying number of VMs. We observed a bottleneck in the network and storage due to increased round trip time (RTT) of the data packets caused by both virtual network layers and CPU context switches (https://en.wikipedia.org/wiki/Contextswitch [5]). We have enhanced virtual network bridge to optimize RTT of data packets by reducing wait time in the network bridge and measured 8, 12% throughput improvement for network and storage respectively. This enhanced network bridge can be used in production with explicit configurations.