On Virtual-Machine-Based Windows File Reads: A Performance Study

Today virtual machine (VM) technologies play an active role in some IT field. But few researches have been done on the I/O performance of the OS running in VM. This paper deals with measurements and analysis of reading large files on Windows systems in VM. Compared with the system on hardware directly, OS in VM owns a more complicate storage hierarchy, which contains the OS cache, VM cache and disk buffer. For this case we implement an evaluation environment where many read tests with different configurations have been completed. The performance study shows an I/O throughput drop by a factor 2 depending on the request size and up to a factor 2 depending on the specified access mode. We evaluate and identify the key issues affecting the performance and a corresponding access model is constructed based on these results. Moreover, a unified VM cache is presented, which can support more than one VM synchronously in a host. Tests show that it can improve the read performance because it avoids storing multi instances of shared data to increase the cache efficiency. Our work can be used to optimize the I/O strategy for a given application or design the VM with higher efficiency.