Barrier enabled QEMU

Abstract

Virtual machines can be run inside one physical machine. This advantage enables cloud service providers and data centers to provide the identical and independent services to multiple users with fewer hardware. QEMU, a virtual machine that is used widely, has the two problems. One is performance degradation due to the amplification of the sync operation. The other is that QEMU not support the BarrierFS which uses the barrier enabled I/O stack that can guarantee the order of writes without transfer- and-flush mechanism. In this paper, we apply the fdatabarrier() system call, which is provided by BarrierFS, to address the I/O performance degradation problem due to the amplified sync operations. We also implement the barrier handler in QEMU so that the QEMU can execute the guest with BarrierFS to reduce the I/O latency caused by transfer-and-flush. The barrier handler recognizes the barrier flag in I/O command from guest and transmits the flag to host storage. We evaluate performance of the QEMU applied BarrierFS compared with the original QEMU. We observe I/O throughput using the varmail of filebench as macro benchmark and Mobibench and GNU dd command as micro benchmark. In result of evaluation, throughput of the QEMU applied BarrierFS is improved by maximum of 13x and minimum of 3.9x compared with the original QEMU. Thus, we verified that the barrier enabled QEMU has higher I/O performance than the original QEMU.