DMVL: An I/O bandwidth dynamic allocation method for virtual networks

In a consolidated server system that uses virtualization, accesses to physical network devices from guest virtual machines (DomUs) need to be coordinated. In this environment, virtualized network devices are required to service workloads executing concurrently from multiple DomUs, with potentially diverse network data delivery requirements. Although a number of methods have been developed for I/O performance virtualization among multiple DomUs, previously proposed researches focused either on improving network I/O performance and lowering overhead from hardware and software, or on achieving network I/O fairness by directly applying special physical network interface cards. We argue that it is important to allocate network I/O bandwidth fairly and stably among multiple DomUs based on pure software approaches and not to hinder live migration and portability of virtual machines. This paper proposes Dynamic Mapping of Virtual Link (DMVL) method, which prevents the interference between multiple DomUs by introducing separated Logical Data Path (LDP) and I/O request queue for each DomU. In DMVL, several techniques are employed. Firstly, we provide isolated I/O bandwidth guarantees to each DomU by mapping a separate LDP to I/O device for each DomU and using multi-queue. Then, the adjusted LDP bandwidth quantity is converted into adjusted credit count, and credit transferring and updating methods based on shared logs are introduced to adjust LDPs bandwidth dynamically. Finally, we improve lottery scheduling algorithm based on shared logs to implement adaptive bandwidth adjustment to meet proportional multi-queue scheduling at lower cost. The proposed techniques are implemented on the Xen virtual network, and evaluated with micro-benchmarks and simulated workloads on Linux guest operating systems. Experimental results show that DMVL improves fairness by at least 60% and stability by at least 29% in the cases of three or more virtual DomUs running on the same physical host.