Virtual Disk Snapshot Management at Scale

Contrary to the other resources such as CPU, memory, and network, for which virtualization is efficiently achieved through direct access, disk virtualization is peculiar. In this paper, we make four contributions. Our first contribution is the characterization of disk utilization in a public large-scale cloud infrastructure. It reveals the presence of long snapshot chains, sometimes composed of up to 1000 files. Our second contribution is to show that long chains lead to performance and memory footprint scalability issues by experimental measurements. Our third contribution is the extension of the Qcow2 format and its driver in Qemu to address the identified scalability challenges. Our fourth contribution is the thorough evaluation of our prototype, called sQemu, demonstrating that it brings significant performance enhancements and memory footprint reduction. For example, it improves the throughput of RocksDB by about 48% compared to vanilla Qemu on a snapshot chain of length 500. The memory overhead on that chain is also reduced by 15x.