Rethink the storage of virtual machine images in clouds

• We propose a three-tier zone based VM image storage model.
• Zone-based model balances requirements of performance, resources, and overhead.
• The zones and caches in Zone-based model have high scalability and availability.
• A friendly VM placement strategy is proposed to improve IO performance of VMs.
• We evaluate some popular models and conclude their advantages and disadvantages.

As one of the most prevalent cloud services, Infrastructure-as-a-Service provides virtual machines (VMs) with high flexibility. How to effectively manage a huge amount of VM images becomes a big challenge. On one hand, images affect VMs’ disk IO performance significantly, which is essential to the quality of services, especially for those having intensive disk IO workloads. On the other hand, they consume many storage resources and cause much management cost, which is cared by cloud managers. Current ways to optimize images usually focus on either improving performance or decreasing image size, which unfortunately cannot satisfy the requirements of high IO performance, low storage consumption, and low management cost simultaneously. Typically, high IO performance requires images storing close to VMs, but this increases redundant data and consumes extra storage at the same time. Besides, a closer image means more data stored in local disks rather than a normal shared storage, which increases management cost as well. In this paper, we analyze these requirements and potential tradeoffs among them, and propose Zone-based model to well balance the requirements. We partition computing nodes into many zones, and construct a shared storage in each zone to cache hot data for high IO performance and low storage consumption. In addition, we improve the normal Copy-on-Write and cache mechanisms, providing new image types and cache functions to enhance the eventual effectiveness. The evaluations show that, our solution improves IO performance by more than 100% in general and even 10 times while adopting a friendly VM placement strategy, and gets close or less storage consumption and management cost than the other models at the same time.