The rapid advances in the development of low-cost computer hardware have led to many proposals for the use of this hardware to improve the performance of database management systems. Usually the design proposals are quite vague about the performance of the system with respect to a given data management application. In this paper we predict the performance of several of the proposed database management machines with respect to several representative INGRES queries. The systems analyzed in this paper include associative disks, RAP, CASSM, DBC, DIRECT, and CAFS. We demonstrate that no one database machine is best for executing all types of queries. We will also show that for one class of queries the degree of performance improvement achieved does not warrant use of a database machine.

Specialized, single function processors can be built to be faster and cheaper than general purpose processors. Most database machines use such special purpose processors to manipulate data, with a general purpose managing processor to control the special purpose processors and perform utility functions. In this paper, the organization and use of these data manipulation processors is explored. Database machines are classified into single data manipulation processor systems, multiple disk-associated data manipulation processor systems, and multiple cache-associated processor systems. Examples of actual database machines are given for each category.Application types are classified into business, bibliographic search, and statistical analysis systems. A metric is developed to compare the performance of the categories of database machines with respect to the application types. The metric is the effective instruction rate, which is comparable to the instruction rate (millions of instructions per second) for conventional computers. It is shown that the effective instruction rate is highly sensitive to the proportion of work performed in the database machine's data manipulation processors. Therefore, determining the work that must be performed in the machine's managing processor is found to be important to determining the performance of the machine.Database machine performance for each category of database machines is compared for each type of application. It is shown that single processor systems are best for business applications; that disk-based multiple processor systems are best for bibliographic search applications; and that hybrid systems are best for statistical analysis applications. Therefore, the design of the organization of data manipulation processors is shown to be application-dependent.

The effect on the performance of data management systems of the use of extended storage devices, multiple processors and prefetching data blocks is analyzed with respect to one system, INGRES. Benchmark query streams, derived from user queries, were run on the INGRES system and their CPU usage and data reference patterns traced. The results show that the performance characteristics of two query types: data-intensive queries and overhead-intensive queries, are so different that it may be difficult to design a single architecture to optimize the performance of both types. It is shown that the random access model of data references holds only for overhead-intensive queries, and then only if references to system catalogs are not considered data references. Significant sequentiality of reference was found in the data-intensive queries. It is shown that back-end data management machines that distribute processing toward the data may be cost effective only for data-intensive queries. It is proposed that the best method of distributing the processing of the overhead-intensive query is through the use of intelligent terminals. A third benchmark set, multi-relation queries, was devised, and proposals are made for taking advantage of the locality of reference which was found.