Network Aware VM Migration in Cloud Data Centers

Abstract

Host virtualization allows data centers to live migrate an entire virtual Machine (VM) to support data center maintenance, disaster avoidance and workload balancing. Live VM Migration can consume nearly the entire bandwidth for memory intensive applications which impacts the performance of competing flows in the network. A network-aware VM Migration operation ensures a fair share allocation of network resources, leading to a seamless Virtual Machine mobility while minimizing degradation of network performance. Recently, VMPatrol was proposed as a network aware VM Migration model which uses a single physical machine and QoS policies to simulate and implement a cost of migration model. However, the performance evaluation of VMPatrol was conducted in an emulated environment. In this paper, we empirically evaluate the performance of VMPatrol in an experimental GENI testbed characterized by wide-area network dynamics and realistic traffic scenarios. We deploy OpenFlow end to end QoS policies to reserve minimum bandwidths required for successful VM Migration. Preliminary results demonstrate that enforcing QoS policies in terms of bandwidth reservation relieves the network of possible overloads during migration. The results indicate that time taken to complete VM Migration depends on VM's memory size, VM page dirty rate and the available bandwidth. The results also indicate that length of stop copy phase and minimum required progress amount are critical parameters in estimating the VM migration cost.