VAGABOND - The Design and Analysis of a Temporal Object Database Management System

Kjetil Nørvåg
Department of Computer and Information Science
Norwegian University of Science and Technology
7491 Trondheim, Norway
noervaag at idi.ntnu.no

Abstract

Storage costs are rapidly decreasing, making it feasible to store larger amounts of data in databases. However, the increase in disk performance is much lower than the increase in memory and CPU performance, and we have an increasing secondary storage access bottleneck. Even though this is not a new situation, the advent of very large main memory has made new storage approaches possible. In most current database systems, data is updated in-place. To support recovery and increase performance, write-ahead logging is used. This logging defers the in-place updates. However, sooner or later, the updates have to be applied to the database. This often results in non-sequential writing of lots of pages, creating a write bottleneck. To avoid this, another approach is to eliminate the database completely, and use a log-only approach, similar to the approach used in log structured file systems. The log is written contiguously to the disk, in a no-overwrite way, in large blocks. This thesis presents the architecture and design of Vagabond, a temporal object database management system (ODBMS) based on the log-only principle. Solutions to problems regarding temporal data management, fast recovery, efficient management of large objects, dynamic reclustering, and dynamic tuning of system parameters are provided. This includes a new index structure for indexing temporal objects, persistent caching of index entries to solve the object indexing bottleneck, algorithms for transaction management, and declustering strategies to be used in a parallel temporal ODBMS. In order to compare the log-only approach with the traditional in-place update approach, analytical cost models are used to study the performance of the approaches. The analysis shows that with the workloads we expect to be typical for future ODBMSs, the log-only approach is highly competitive with the traditional in-place update approach. Many of the ideas presented in this thesis are also useful outside the log-only context. In papers included as appendixes, we show how the ideas can be applied to temporal ODBMSs based on traditional in-place updating techniques.

The whole thesis is available as postscript, gzip'ed postscript, and pdf.
Last modified: Fri May 24 19:05:20 MEST 2002