Performance Evaluation of Non-tunneling Edge-Overlay Model on 40GbE Environment

Network virtualization is an essential technology for cloud datacenters that provide multi-tenancy services. Cutting-edge SDN-enabled datacenters have introduced an edge-overlay (distributed tunneling) model to construct virtual tenant networks. The edge-overlay model generally uses L2-in-L3 tunneling protocols, and VXLAN is a representative of such protocols. However, the tunneling-based edge-overlay model has some performance and compatibility problems. We have proposed yet another edge-overlay model that does not rely on L2-in-L3 tunneling protocols so far. Our model utilizes two methods, OpenFlow-based address translation and host-based VLAN. The former method replaces VMs' MAC addresses to physical servers' ones, which prevents unnecessary IP fragmentation as well as unnecessary MAC address learning by physical switches. The later method breaks a limitation of the number of VLAN-based virtual tenant networks (4094) by allocating entire VLAN ID space to each physical server and by mapping VLAN ID to VM with OpenFlow controller support. Our previous study has shown that performance of the proposed model is superior to existing other models on 1GbE environment. However, the upper bound of our proposed method, performance behavior with large packet size, and effect of hardware-assisted technology like segmentation offload have not been evaluated. In this paper, we evaluate performance effects of these aspects on high-bandwidth (40GbE) environment. The results show that the performance of VM-to-VM communication with the proposed model is close to that of physical communication and exceeds 10Gbps throughput with large TCP segment.