Caching VM instances for fast VM provisioning: A comparative evaluation

Abstract

One of the key metrics of performance in an infrastructure cloud is the speed of provisioning a virtual machine (or a virtual appliance) on request. A VM is instantiated from an image file stored in the image repository. Since the image files are large, often GigaBytes in size, transfer of the file from the repository to a compute node running the hypervisor can take time in the order of minutes. In addition to it, booting an image file can be a time consuming process if several applications are pre-installed. Use of caching to pre-fetch items that may be requested in future is known to reduce service latency. In order to overcome the delays in transfer and booting time, we prepare a VM a priori, and save it in a standby state in a "cache" space collocated with the compute nodes. On receiving a matching request, the VM from the cache is instantly served to the user, thereby reducing service time. In this paper, we compare multiple approaches for pre-provisioning and evaluate their benefits. Based on usage data collected from an enterprise cloud, and through simulation, we show that a reduction of 60% in service time is achievable. © 2012 Springer-Verlag.