I/O Performance Analysis And Optimization In Virtualized Environments

Since 21st Century,we have more cloud practitioners,better cloud ecosystem,and more companies embrace the cloud platform,which give bigger challenge to the virtualization technologies.I/O virtualization is an important part of virtualization technologies and is the hot topic for academic and industry.The key problem of I/O virtualization is to make data storage faster,persistent and consistent.For performance consideration,we usually add a cache layer to the system.However,we would lose data if the system crashed.To make data persistent and consistent,the system needs to call the expensive sync operation,which slows the system down.This paper deeply analyzes the whole process of I/O virtualization and proposes several optimizations to improve the I/O performance.The main contributions of this paper are divided into three parts:1.In this paper,we analyze the virtual disk format and I/O process of virtualization,and find out the sync amplification problem and why the problem happens.The sync amplification problem means that the virtual disk will introduce more sync operations than the guest(e.g.,more than 3X for QCOW2 and more than 4X fore VMDK).The reason for the sync amplification problem is that the virtual disks introduce many metadata for the image management.Such metadata requires extra sync to keep the consistency during the modification of virtual disks.2.To solve the sync amplification problem,this paper proposed three optimizations,namely dual-mode journal,dynamic pre-allocation and NVRAM log.These optimizations improve the I/O performance as well as keep the consistency of virtual disks.Our experiments show that the three optimizations improve the 110%performance of varmail workload and improve the 50%performance of TPCC workload.3.This paper implemented the external-sync system in the virtualization environment.From the perspective of the user,external-sync system commits several synchronous I/O in a group,which reduces the number of sync operation and improves the I/O performance.Our tests show that the external-sync system improves the I/O performance by 46 times under the sequential synchronous write workload,and the external-sync system improves the I/O performance by 50%under the TPCC workload.