The optimization of cost-model for join operator on spark SQL platform

Abstract

Spark needs to use lots of memory resources, network resources and disk I/O resources when Spark SQL execute Join operation. The Join operation will greatly affect the performance of Spark SQL. How to improve the Join operation performance become an urgent problem. Spark SQL use Catalyst as query optimizer in the latest release. Catalyst query optimizer both implement the rule-based optimize strategy (RBO) and cost-based optimize strategy (CBO). There are some problems with the Catalyst CBO module. In the first place, the characteristic of In-memory computing in Spark was not fully considered. In the second place, the cost estimation of network transfer and disk I/O is insufficient. To solve these problems and improve the performance of Spark SQL. In this study, we proposed a cost estimation model for Join operator which take the cost from four aspects: Time complexity, space complexity, network transfer and disk I/O. Then, the most cost-efficiency plan could be selected by using hierarchical analysis method from the equivalence physical plans which generated by Spark SQL. The experimental results show that the total amount of network transmission is reduced and the usage of processor is increased. Thus the performance of Spark SQL has improved.