HV2M: A novel approach to boost inter-VM network performance for Xen-based HVMs

• We analysis the key issues in inter-HVM network performance on Xen.
• We propose a novel hardware-independent approach which pushes data like an express to boost inter-HVM network performance.
• Our new approach is designed to use less shared memory but can achieve higher performance.
• We design and implement a prototype system to demonstrate our idea.
• Through comparison with PV-on-HVM and SR-IOV, evaluation shows our prototype is in big advance on network throughput.

Despite the rapid development and the wide use, virtualization confronts new challenges on improving network I/O performance. For virtual machines co-existed on a server, inter-VM network performance turns out to be worse than expected. Although many efforts have been dedicated to that issue on paravirtualized (PV) machines, HVMs (Hardware Virtual Machines) get less attention. However, co-resident HVMs are also suffering from low network I/O performance which troubles IaaS cloud infrastructure providers. Such problem is prominent as in the front-back network model several performance bottlenecks appear when inter-VM traffic is being handled.In this paper, we propose a novel approach and implement a prototype called “HV2M” to optimize inter-HVM network performance, which mainly address the throughput issues. HV2M takes advantages of the existing mechanisms for HVMs provided by Xen and is totally hardware independent. Unlike the existing idea, HV2M pushes data to the receivers without using shared memory buffer to transfer the pending network data. To evaluate our work, we test and compare network response time, network throughput and CPU utilization with other popular approaches such as SR-IOV. The result shows that the new communication model has great improvements in network delay and throughput with little pressure on Domain-0.