Active Disk Architecture for Databases (CMU-CS-00-145)

Today's commodity disk drives, the basic unit of storage for computer systems large and small, are actually small computers, with a processor, memory and a network connection, in addition to the spinning magnetic material that stores the data. Large collections of data are becoming larger, and people are beginning to analyze, rather than simply store-and-forget, these masses of data. At the same time, advances in I/O performance have lagged the rapid development of commodity processor and memory technology. This paper describes the use of Active Disks to take advantage of the processing power on individual disk drives to run a carefully chosen portion of a relational database system. Moving a portion of the database processing to execute directly at the disk drives improves performance by: 1) dramatically reducing data traffic; and 2) exploiting the parallelism in large storage systems. It provides a new point of leverage to overcome the I/O bottleneck. This paper discusses how to map all the basic database operations - select, project, and join - onto an Active Disk system. The changes required are small and the performance gains are dramatic. A prototype based on the Postgres database system demonstrates a factor of 2x performance improvement on a small system using a portion of the TPC-D decision support benchmark, with the promise of larger improvements in more realistically-sized systems.