人工神经网络预测并行查询响应时间

Predicting response time of parallel queries with artificial neural network

针对现有查询响应时间预测统计模型存在准确率无法提高、特征选取单一、动态性差的问题,综合考虑查询计划、查询交互两大因素,提出采用结构简单、易搭建的人工神经网络——全连接神经网络预测并行查询响应时间。采集查询计划与查询交互数据作为输入特征,查询真实的响应时间作为预测标签,训练模型,进行预测。此方法不需要预先知道样本数据的数学模型函数,仅通过对样本数据集的学习建立模型,建模过程简单,可达较好的预测效果。实验结果表明,全连接神经网络模型准确率高达79.99%,较当前代表性的统计模型提高约6%。

To resolve the problems of existing statistical models that the accuracy rate cannot be improved,the feature selection is single,and the dynamicity is poor,the fully connected neural network was used for predicting the response time of parallel queries through considering the influences of query plan and query interaction.A concise and easy neural network possessed was used to build structure.The data of query plan and query interaction were collected as input features,and the practical response time of queries was used as the prediction label.The fully connected neural network could get trained and yielded models for predicting.The method dis not require the mathematical model function of sample data in advance and could achieve a better prediction performance with a simple modeling process,models were learned and trained for collected sample dataset.Experimental results show that the model proposed achieves an accuracy with 79.99%,which performs roughly 6%higher than the state-of-the-art statistical model.