Designing Transactional Hypervisor for Improving the Performance of Qcow2-Based Virtual Disks

Abstract

Virtualization has become one of the backbone technologies for constructing cloud servers and data centers, in that the technology supports to partition the physical computing resources of the system into multiple logical ones for concurrently running multiple VMs (virtual machines) in a single physical server. In constructing multiple VM environments, the qcow2 format is widely used as a virtual disk layout due to its attractive features such as high storage space efficiency and high reliability. Unfortunately, according to the policy of the qcow2 format, the hypervisor calls file sync operations excessively during allocating data clusters. Moreover, significant performance degradation may occur due to the CoW (Copyon-Write) operations that are unconditionally performed by the hypervisor whenever new data clusters are allocated on the virtual disk. In this paper, we introduce a transactional hypervisor that mitigates the performance overhead caused by the data cluster allocation onto the qcow2-based virtual disk. To achieve this, we adopt transactional support to the existing hypervisor and propose a new type of file sync operation, called gsync, which flushes the modified metadata in a bundle. Moreover, we introduce an optimization technique on the CoW mechanism that performs CoW operations selectively according to the amount of data that will be updated on each data cluster. The experimental results clearly show that the proposed hypervisor outperforms the conventional hypervisor in terms of write IOPS (Input/Output Per Second) by up to 78.4%, by reducing the number of file sync calls and that of CoW operations by up to 50.6% and 52.6%, respectively.