Selecting Execution Plan For Concurrent Queries Using LSTM-FCN

Query is the main load of database system,and the efficiency of query execution determines the performance of database system.Choosing an appropriate execution plan for query is the key to improve the performance of database,and eventually that of the application system.At present,the query optimizer is mainly based on the cost model to select the execution plan for the query.The establishment of the cost model usually depends on the statistical information in the database system,and the accuracy of the statistical information is affected by the data distribution.Also the statistical information provided by the optimizer deviates greatly from the real situation if the data distribution is unknow.For concurrent queries,the state-of-the-art query optimizers select a “not bad” plan for concurrent queries by configuring parameters such as parallelism.Since the parameters are fixed,plans selected by the optimizers are also fixed.In fact,there exists query interaction(QI)among concurrent queries in database systems.QI usually causes the change of the amount of resources or time used by the query,or cost drift.QI varies in different query mixes,and cost drift is also different.Plan selection through configuration parameters often fails to reflect this change.Therefore,QI should be taken into account to select an “appropriate” plan for concurrent queries dynamically.There are numerous query mixes and their QI is too complex and dynamic to model with traditional modelling approaches.This paper studies the approach by using deep learning neural networks to select a plan for concurrent queries.Considering that relational operators of a plan are evaluated in sequence,this paper proposed a concurrent query execution plan selection method based on LSTM-FCN using the time domain characteristics of LSTM(Long Short-Term Memory)network,and the advantages of FCN(Full Connected Networks)for feature fusion and classification.Design and code the execution plan and interaction characteristics for the query mix,then feed them into the network,thereby selecting an appropriate execution plan for the query in actual operating scenarios dynamically.Experiments on Postgre SQL verify that the method proposed in this article is feasible and effective.LSTM-FCN selects the appropriate execution plan based on the average accuracy of 97.1% for the query under the conditions of different query mixes and parallelism of 3,4,5,6,and 7.