THE USE OF IMMENSE MEMORIES FOR DATABASE-SYSTEMS - A PRAGMATIC APPROACH

We examine the cost and performance advantages of immense main memorie s and of the use of extended storage units for a wide range of databas e applications. Most database systems are required to guarantee a clos e to absolute degree of data integrity. This requirement is a serious limitation on the usage of main memories as the primary repository of data because of the vulnerability of the content to errors, particular ly, software related errors. Current research often neglects this poin t and focuses instead on the technological issues of the silicon memor ies content volatility. Transaction processing performance is mostly e valuated on the basis of throughput (i.e., transaction per second) pro vided that the average response time does not exceed 1 s. Most update intensive applications can be classified as transaction processing and therefore driven by throughput considerations. We will show that larg e memories have little effect on throughput. Given this point of view, there is very little (if any) performance advantage in using main mem ories for transaction processing. Consequently, the potential of the i mmense memory system is for improving the performance of those (non-up dating) applications that execute complex queries. The coexistence of query applications with the updating transactions is a realistic view of most business systems as well as many engineering and scientific ap plications. By taking these facts into consideration, a new vision of in-memory databases is created. In this approach, hardware architectur e changes are shown to be less important than the use of proper algori thms and software systems. (C) 1998 Elsevier Science Inc. All rights r eserved.