Fine-grained Autoscaling with In-VM Containers and VM Introspection

Abstract

Clouds often provides a mechanism called autoscaling to deal with load increases of services running in virtual machines (VMs). When a VM is overloaded, scale-out is performed and automatically increases the number of VMs. However, when multiple services run in one VM, the entire VM is always scaled out even if only one service is over-utilized. In this case, only an over-utilized service should be scaled out, but it is not easy for clouds to accurately monitor the resource usage of services inside VMs. This paper proposes Ciel, which runs each service in a container created inside a VM for separation of services and enables fine-grained autoscaling of VMs. Using VM introspection, Ciel accurately monitors the resource usage of each in-VM container from the outside of a VM in a non-intrusive manner. If it detects an overloaded in-VM container, it creates a new VM of minimum size and boots only the container that needs to be scaled out in the VM. This can minimize both the cost of the VM and the time taken for scale-out. We have implemented Ciel using Xen and Docker and showed the effectiveness.