Experimental study of virtual machine migration in support of reservation of cluster resources

TLDR

Virtual Machines enable efficient resource consolidation and secure application environments, but migrating multiple VMs is time‑consuming and requires estimation. The study aims to model and predict VM migration performance to support CPU, memory, and network reservation for individual VMs and clusters. The authors experimentally analyze VM migration in scenarios where physical servers are vacated to launch VM clusters for parallel jobs, developing a predictive model of migration performance. The experiments demonstrate that migration times can be predicted for different VM configurations, that parallel migration shortens total time yet increases per‑VM latency, and that buffering VM state in main memory improves performance without disk writes.

Abstract

Virtual Machines are becoming increasingly valuable to resource consolidation and management, providing efficient and secure resource containers, along with desired application execution environments. This paper focuses on the VM-based resource reservation problem, that is, the reservations of CPU, memory and network resources for individual VM instances, as well as for VM clusters. In particular, it considers the scenario where one or several physical servers need to be vacated to start a cluster of VMs for dedicated execution of parallel jobs. VMs provide a primitive for transparently vacating workloads through migration; however, the process of migrating several VMs can be time-consuming and needs to be estimated. To achieve this goal, this paper seeks to provide a model that can characterize the VM migration process and predict its performance, based on a comprehensive experimental analysis. The results show that, given a certain VM's migration time, it is feasible to predict the time for a VM with other configurations, as well as the time for migrating a number of VMs. The paper also shows that migration of VMs in parallel results in shorter aggregate migration times, but with higher per-VM migration latencies. Experimental results also quantify the benefits of buffering the state of migrated VMs in main memory without committing to hard disks.