Design, implementation, and benchmarking of a file manager for a relational database management system on a raw UNIX disk

Abstract

The file management component of a database management system (DBMS) has to be tailor designed to meet the performance demands of large database applications. The operating system (OS) file systems are typically not suitable for storing and manipulating large DBMS files. First, performance is important but not a major issue in the design of OS files. Second, the DBMS buffer manager with a priori information is more sophisticated than the OS counterpart. Finally, the OS file system does not provide physical clustering of related files, a feature highly desirable in a DBMS environment. On the other hand, a DBMS is typically characterized by large file sizes and sequential accesses. Therefore, a DBMS on top of an OS file system can result in considerable disk arm movement and hence large seek times, adversely affecting the application performance. Database technology has matured and is rapidly shifting toward the distributed and client-server systems. The number of applications for it has also grown. These new applications demand more sophisticated data access and handling requirements. Modstar is a file manager on a raw UNIX disk that will bypass the UNIX OS file system. The implementation on a raw device allows us to manage our own buffers without the additional overhead of copying to and from the UNIX buffers. Modstar is designed for database applications. The experimental Modstar is an insight into the internals of a DBMS, and in understanding issues such as the impact of storage structures, join algorithms, database updates, etc., on the performance of query processing.