Hypervisor and Neighbors’ Noise: Performance Degradation in Virtualized Environments

Users expect isolated performance from rented virtual machines (VMs) in an infrastructure as a service (IaaS) cloud environment. However, this is not happening in todays’ systems because basically VMs are running in a shared environment. In this paper, we study performance degradation in a virtualized environment similar to IaaS clouds using Parsec 2.1 benchmarks. We consider slowdowns caused by hypervisor—hypervisor's noise—as well as co-located VMs—neighbors’ noise. Previous researches did not consider multi-virtual CPU (vCPU) VMs in an overcommitted environments similar to IaaS clouds. Our target system consists of multiple multi-processor VMs running on a commodity chip-multiprocessor by a hypervisor. This configuration is widespread in todays’ IaaS clouds like Amazon EC2. We find that performance degradation in a virtualized environment could be up to $16\times$ which is far more than previous findings. Beside shared resources of memory sub-system, blindness of hypervisor's scheduler have large impact on the slowdown and this is contrary to recent researches that mostly blame last-level cache (LLC) contention for performance degradation. After investigating the causes of performance degradation, we provide some ideas that motivate researchers to reduce performance degradation through hardware and software techniques. We also mention some hints that help organizations to see if their applications are ready for the cloud.