Analyzing and Improving Table Space Allocation

Abstract

Space allocation is a fundamental operation performed by a database management system (DBMS) when it inserts a record into a table. A good space allocation algorithm quickly locates and reserves enough space for a record, places it closer to its related records, and utilizes the available space. Satisfying these conflicting requirements is challenging and trade-offs are carefully balanced by wellchosen heuristics. As a DBMS evolves over time, especially a commercial DBMS, its space allocation algorithm gets more sophisticated and complex and relies on many heuristics. Technological changes, new applications, and greater data volumes render many legacy heuristics ineffective. These factors hinder understanding of space allocation behavior under many workload conditions and make it difficult to enhance the algorithm without causing performance regressions for some of the workloads. To facilitate research and study the performance of a table space allocation algorithm of a modern DBMS in real-world workload scenarios, we build an extensible simulation framework. We analyze algorithm behavior and make surprising observations. We use the findings to further improve the existing algorithm by proposing algorithm enhancements and showing their benefits with respect to key performance metrics. In conclusion, the proposed framework has been effective in research to understand the performance, improve the space allocation algorithms, and to guide the developers of a commercial DBMS.