Multiported storage devices

Abstract

In the past decade the demand for systems that can process and deliver massive amounts of storage has increased. Traditionally, large disk farms have been deployed by connecting several disks to a single server. A problem with this configuration is that a given storage server can become a bottleneck, as all disk activity for a particular server has to be processed by that server. Increasing performance and decreasing cost of microprocessors are making it feasible to move more processing power to the data source. This allows us to investigate new methods of storage delivery that were not plausible in the past. In this thesis we present a new "enhanced" storage device called a Multiported Storage Device which is endowed with more processing power and intelligence than the traditional block storage device. A multiported storage device allows application-specific code that we call filter applets to be downloaded to the device while still maintaining the simple block-level interface. The device contains several software ports through which read and write requests pass. Each filter applet is associated with a particular software port in the device. We present a prototype implementation of a multiported storage device using the Linux operating system. We implement our device using kernel driver modules and by performing several modifications to the Linux kernel to accommodate the multiported storage device architecture. We also present experimental results showing the impact of our layered drivers on I/O performance.