Design of the 64-bit Option for the Oracle7 Relational Database Management System
Vipin V. Gokhale · Digital technical journal · 1996
Historically, the limiting factor for the Oracle7 relational database management system (RDBMS) performance on any given platform has been the amount of computational and I/O resources available on a single node. Although CPUs have become faster by an order of magnitude over the last several years, I/O speeds have not improved commensurately. For instance, the Alpha CPU clock speed alone has increased four times since its introduction; during the same time period, disk access times have improved by a factor of two at best. The overall throughput of database software is critically dependent on the speed of access to data. To overcome the I/O speed limitation and to maximize performance, the standard Oracle7 database server already utilizes and is optimized for various parallelization techniques in software (e.g., intelligent caching, data prefetching, and parallel query execution) and in hardware (e.g., symmetric multiprocessing [SMP] systems, clusters, and massively parallel processing [MPP] systems). Given the disparity in latency for data access between memory (a few tens of nanoseconds) and disk (a few milliseconds), a common technique for maximizing performance is to minimize disk I/O. Our project originated as an investigation into possible additional performance improvements in the Oracle7 database server in the context of increased memory addressability and execution speed provided by the AlphaServer and DIGITAL UNIX system. Work done as part of this project subsequently became the foundation for product development of the Oracle 64 Bit Option. Of the memory resource that the Oracle7 database uses, the largest portion is used to cache the most frequently used data blocks. With hardware and operating system support for 64-bit memory addresses, new possibilities have opened up for high-performance application software to take advantage of large memory configurations. Two of the concepts utilized are hardly new in database development, i.e., improving database server performance by caching more data in memory and improving I/O subsystem throughput by increasing data transfer sizes. However, various conflicting factors contribute to the practical upper bounds on