Distributed indices for accessing distributed data

Abstract

Distributed-search structures are being developed to address the problem of indexing distributed data. The design goal is to create distributed indices that are highly parallel, require low overhead, and can perform data load balancing. These structures can be used to provide highly parallel and scalable indexing on a multiprocessor, and a versatile and low-overhead dynamic index on a distributed system. The first distributed-search structure considered was the dB-tree, a distributed B-tree. It was shown how a highly-parallel but low-overhead dB-tree can be implemented by carefully controlling the degree of replication, and by using lazy updates. A significant use of a distributed index is to aid in load balancing. It was shown experimentally that simple heuristics using out-of-data load information can perform very effective load balancing.<>