A Study of the Performance of Diskless Web Servers

One commonly held assumption is that one or more high performance local disks are required for good server performance. This report examines this assumption for one important category of servers, web servers, by comparing their performance in three different configurations with all files on a local disk, with the content served accessed from a remote machine using NFS and with all files, content and system files, accessed using NFS. Based on an examination of the performance of three static web-serving workloads constructed from traces on typical, commodity hardware, we conclude
that the location of the content makes little difference in the responsiveness. Additionally, using fully diskless operation with all files, both system and content, accessed over NFS makes very little difference in responsiveness. To stress the remote file access subsystem, we increased the intensity of the workloads to the point at which response time became unacceptable. We found that by the time that it became unacceptable for remote file access, it was also unacceptable for local file access. The
differences in the points of degradation that are observed may be attributable to memory overload in the remote access case, particularly in the case which uses no local disks at all. However, verfication of this hypothesis is beyond the scope of this report. Based on this evaluation, we conclude that any performance loss experienced by diskless web servers is outweighted by their other benefits, such as their simpler administration, lower power consumption and lower heat production.