Pluzhnikov

Connection of two tables in parallel database system (PDBS) was analyzed in this article. The Laplace–Stieltjes transform (LST) of connection time of two tables with NLJ and HJ methods in PDBS was introduced. Random character of the number of rows in the connecting tables is considered along with node fragmentation of tables, filtration of all table entries, probability of matching of connection attribute values, variants of hash-table location in the memory and parameters of interprocessor communication of PDBS nodes. Generating function of resultant value of entries, obtained after connection of tables in any node is presented in the article.

 The authors propose an approach to estimation of time characteristics of query execution to a parallel database system (PDBS). A model of query execution to PDBS with a p F(R) plan as a closed-loop queue system was developed. It was shown that the model could be reduced to a “repairman” model if there was a bottleneck. The possibility of reducing this model to an open-loop queue system is proved. Parameters of this model are determined. A policy of calculation of performance indexes of PDBS built on any architecture (SE, SD, SN) is formulated.

In the article the authors suggest expressions which make it possible to estimate the aggregate value of possessing database parallel systems (DPS). The authors developed an algorithm of DPS selection which allows for specific character of the comparison of DPS architectures and special feature of cost estimates. By means of sequential growth of the number of AMP-processors and SMP-units the algorithm orders the database parallel systems with SE, CE and SN architectures according to their cost increase. It is shown that in this sequence it is possible to isolate the subsequence in which DPS are ordered according to descending ordering of the mean time of responding to requests. As in the algorithm all DPS versions are consecutively analyzed, in the finite number of iterations optimal solution will be found

This article analyzes existing architectures of parallel database systems. The process of query execution in parallel database systems is reviewed in details. Laplas-Stiltes transform is produced for time of query execution in parallel database systems. Properties of Laplas-Stiltes transform are used for deriving of averages of distribution of query execution time for different parallel database systems architectures. Practical example of query execution time computation for SE, SD and SN architectures is shown. Dependence of query execution time from number of processors in defined architecture is analyzed and corresponding conclusions are made.