Virtual Machine Images as Structured Data: The Mirage Image Library

The rise of infrastructure-as-a-service (IaaS) clouds, both private and public, has created new problems in managing large collections of virtual-machine (VM) images. VM images must be kept up-to-date with security patches and scanned for malicious or improperly licensed software. Because images are bulky, attention must be paid to the latencies of deployment and capture. Large collections of VM images also present new opportunities which, while not completely realized today, do not exist with collections of physical machines. Users can search for images that contain the software they want, configured as they want it. Collections can be mined for patterns in order to answer questions like “which database management systems does our company use?” Images can be compared with one another. This paper describes Mirage, an image library that addresses these problems and opportunities and is pluggable into various clouds. Mirage stores images in a format that indexes their filesystem structure, instead of as opaque disk images. Like other libraries, Mirage provides features for image capture and deployment. In addition, Mirage maintains a provenance tree that records how each image was derived from other images; allows operations, like patching and scanning, that normally require a VM instance to execute offline; and enables analyses such as image search and comparison. Other papers have used Mirage [16, 13, 17] or described parts of the system [8]. This paper is the first overview of the complete system. It also describes three novel features that reduce the costs of translating between disk images and the Mirage format: a content-addressed store optimized for VM image data; a simple yet flexible indexer that supports a variety of filesystem and image types; and delta deployment, which uses precomputed deltas between images to speed up format conversions. Finally, we relate our experience with Mirage in the IBM Workload Deployer product (IWD) [3], which serves images to a customer’s own private cloud, and in the Research Compute Cloud (RC2) [10], which is a production IaaS cloud.