An Analysis of Disk Performance in VMware ESX Server Virtual Machines

The performance of applications running within VMs is a significant factor in their adoption. VMware ESX Server was designed for high performance, and its architecture is streamlined to provide high-speed I/O. In this paper, we focus on one component of ESX Server's I/O architecture, its storage subsystem. We look at the characteristics of a series of disk microbenchmarks on several different storage systems: a direct-attached disk, a RAID array and a storage-area network (SAN). We compare the throughput obtained on the native server to the throughput on a virtual machine. We show that despite differences in the I/O characteristics of the different storage subsystems, the behavior of the VM matches that of the native server closely; for example, in the configurations where the native server shows higher throughput, the virtual machine shows a similar increased throughput. Next, we discuss how to apply the microbenchmarks to model virtual machine performance for disk-intensive applications, and present case studies for two workloads: a simple file server and a commercial mail server. We show that in these cases, the microbenchmark data, in conjunction with characteristics of the native application, can be used to generate an accurate estimate of the application’s performance in a virtual machine. VMware ESX Server is a software platform that efficiently multiplexes the hardware resources of a server among virtual machines. This paper studies the performance of a key component of the ESX Server architecture: its storage subsystem. We characterize the performance of native systems and virtual machines using a series of disk microbenchmarks on several different storage systems. We show that the virtual machines perform well compared to native, and that the I/O behavior of virtual machines closely matches that of the native server. We then discuss how the microbenchmarks can be used to estimate virtual machine performance for disk-intensive applications by studying two workloads: a simple file server and a commercial mail server.