I/O Congestion-Aware Computing Resource Assignment and Scheduling in Virtualized Cloud Environments

Abstract

Recently, cloud computing based on virtualization technology has developed rapidly. A cloud computing provider offers services mainly by renting virtual machine (VM) to users, through which users can share various computing, storage and network resources. However, VM consolidation in cloud computing environments leads to significant degradation of the network performance of VMs, such as a huge drop in throughput and simultaneously a severe increase in latency and jitter. In this paper, we provide a deep insight into the relationship between VCPU scheduling and I/O congestion in Xen, particularly in data receiving direction. Through experiments, we observe that the VCPU scheduling plays a key role in the network performance of VMs due to the congestion in I/O ring. Then we propose a novel mechanism called Diff-Scheduler. It can identify the so-called NSVM (Network Queuing Sensitive Virtual Machine) that is more vulnerable to the congestion in I/O ring. Furthermore, it can schedule the VCPUs of the NSVMs more frequently than other VMs in different pools independently, without violating the CPU share fairness in VM consolidation. Evaluations based on a prototype of Xen platform featured Diff-Scheduler show that our mechanism significantly improves the network performance of VMs. Specifically, Diff-Scheduler proposed in this paper jointly enhances throughput, latency, and fairness of resource allocation.