A multiprocessor computer architecture for database support

Abstract

The effects of certain computer architecture design principles upon the design of a back-end database machine architecture have been assessed. The example computer architecture, referred to as REPT (RElational Processing Tree), was designed using the principles of multiple identical processing nodes, minimization of global knowledge, and ease of processing expansion. A binary tree communication topology was investigated for the implementation of relational algebra operations. An indirect search of the database was employed by distributing source relations across leaf node memories. Implementations of complex relational algebra expressions were discussed. Current database computer architectures were surveyed. Response time expressions for a representative query were developed for REPT and for each surveyed machine. The results of this analysis were compared using specific examples of data dependent parameter values. The response time analysis showed a general ranking of currently proposed database machines. Problems associated with large result sets were exhibited in this analysis, and in other cases, the expansion of the number of processing cells reduced the response time. A machine which directly searched the database was found to have the fastest response time when the number of result tuples was small. REPT supported a high degree of parallel processing across leaf nodes and pipelining of source and result relations. It was determined that X-NODES (Sequin et al. Communication in X-TREE, a Modular System. Proceedings of the ACM 1978:130-140) could support the desired capabilities of REPT. Outside of the functions performed during query execution, all node processors were identical. REPT was found to be easily expanded. In terms of response time, REPT was found to be competitive with other database machines in its class. Evidence supported the conclusion that the design principles used for REPT are worthy of future computer architectures.