Research On Query Push-down Strategies For A Shared Storage Distributed Database

In a shared storage distributed database where computation and storage are separated,all computation is completed at the computing layer.If the tables or intermediate result sets involved in computation are very large,it may occur that due to the lack of memory in the computing layer,the data may be I/O multiple times between the computing layer and the storage layer,which seriously affects the query efficiency.In order to reduce the data transfer between the computing layer and the storage layer,the push-down strategy of distributed database query with shared storage is studied.The scenario suitable for query push-down is analyzed,and the query cost estimation model with query push-down and the query rewriting strategy based on cost estimation are given for WHERE clause filter condition,group query,join query and other types of query statements.The query push-down strategy is implemented on a distributed database prototype with shared storage,including query cost estimation module and query push-down plan generation module.The query optimizer is modified to support query push-down cost estimation,and the original query execution path is changed to serial/parallel execution mode that can be pushed down to the storage layer.The query push-down module is implemented,including the query push-down control module,the query push-down plan sending module,the query push-down execution module and the query push-down result aggregation module.Ensure that the push-down query task completes and the compute node returns the complete result to the client.For different types of query statements,the function verification and performance comparison of query push-down policy are experimented.The experimental results show that an appropriate query push-down strategy can improve query efficiency and reduce network overhead effectively.The query cost estimation model with query push down can effectively select the efficient query execution path,and the query performance is significantly improved when the storage nodes execute in parallel.