Efficiency of text scanning in bibliographic databases using microprocessor-based, multiprocessor networks

Abstract

Boyer-Moore pattern matching algorithm was used to search 35 natural language quenes against a file of 1000 titles and abstracts taken from the Library and Information Science Abstracts database. A model of searching using a singly-linked chain containing up to 11 transputers was carried out: the maximum speed-up obtained with this size of network was 10.4 with a processor utilization of 0.95, both figures being close to the ideal of 11.0 and 1.0. Expenments with a nearest neighbour searching algorithm for serial document files demonstrate the need to keep the processors fully occupied with computational work if a high degree of speed-up is to be obtained.

Parallel computing in information retrieval – an updated review

“…Parallelism can be used in both phases of the search. Two-phase searches have been implemented on SIMD machines by Pogue et al [11][12][13][14][15] and on a MIMD transputer network by Cringean et al [58][59][60][61][62]. Panagopoulos and Faloutsos [55] also recommended the method's use when using signature files.…”

“…An example of the two-phase search can best be illustrated by looking at one particular system, the transputer network program described by Cringean et al [58][59][60][61][62]. This system uses the process farm approach to parallelism to increase efficiency on the more computationally intensive second phase.…”

“…One of the main methods used to improve retrieval effectiveness is to utilise one of the myriad term weightings schemes that are available. The dominant scheme has been the vector processing model with systems such as RAP.3 [75], the DowQuest [63], Transputer Networks [58][59][60][61][62]84] and POOMA [68], all using it in various forms. PADRE [20] offers a number of weighting schemes based on the inverse document frequency (IDF) measure.…”

“…These models may use unnormalised [20] or normalised [24,63,68,84] term weighting. Cringean et al [58][59][60][61][62] do not specify the normalisation method. Others such as Reddaway [11] and Jeong and Omiecinski [69] do not specify a weighting scheme in their discussions of term scoring in their papers.…”

Atom-by-atom searching using massive parallelism. Implementation of the Ullmann subgraph isomorphism algorithm on the distributed array processor

“…The distribution of the database may be done statically, by assigning each PE one record (or group of records), with all PEs sharing the same code; a review of the use of processor arrays for database searching using this approach is provided by Willett and Rasmussen.27 Alternatively, the distribution of the database can be carried out dynamically by means of load balancing procedures, which ensure that records are assigned to PEs during program execution as the latter become available. This model of computation is often used with parallel machines that make use of the processor farm model of computation, where computational work is allocated to processors as they become free for further processing; the use of processor farms for database searching is discussed by Cringean et al 35 and by Downs et al11 These two, rather different approaches to parallel searching are exemplified by the two DAP implementations of the Ullmann algorithm that we have studied as described below. 36 Algorithm I uses inner loop parallelism while Algorithm II uses outer loop parallelism.…”