Abstract

This paper presents some experimental results on the performance of distrihuted join algorithms in a local nctwol-k. Eight different join methods have been imple-mented in an experimental distrihuted system, the Cry-seal multicomputcr, and tested for join queries with a variety of relation sizes, join selcctivitics, and join column value distributions. The results obtained indi-cate that pipelined join methods outperform sequential methods over a wide range of join queries. II was also found that the communications costs in a local network environment arc not a dominant factor with rcspcct IO performance, and that shipping a whole relation from one site to another is not an unreasonahlc dccison as long as it is done in the framework 01 ‘ pipelined algo-rithms. l’wo pipelincd variants 01 ‘ a centralized nested loops join algorithm (with an index) wet-c found to per-form well lor many oi ~hc queries tested. 1. INTRODI’CTION In relational datahasc systems, qucrics arc typically posed in a high-lcvcl, nonprocedural query language hascd on the relational calculus such as QUEL or SQL [UllmX2]. It is the task of Ihc query optimizer to decompose the query into primitive relational operations such as selection, prqjcction and join. The join opera-tor has attracted a great deal of research intcfest since the costs of availahlc join methods tend to vary widely

Citations