Abstract: To address the challenges of continuous growth of data volume,the diversification and complexity of query workloads,a hybrid approach of Extreme Learning Machine(ELM)with Differential Evolution(DE)named DE-ELM was presented for concurrent query performance prediction,in which ELM was used to predict concurrent query performance and DE was adapted to search for optimal network structure and feature subset synchronously.DE-ELM only used information available at compile time,and did not need to specify the number of features in advance,nor need to make prior assumptions on the nature of query interactions or the internal mechanism of database system.Experimental evaluations were executed on top of real-life and TPC-DS benchmark with dynamic concurrent workloads to investigate the effect of simultaneous feature selection and network structure optimization.The results showed that the average accuracy of DE-ELM was over 80%,which verified the feasibility and effectiveness of the proposed method to some degree.

Key words: concurrent query, performance prediction, extreme learning machine, differential evolution, feature selection

摘要： 为应对数据规模持续增长、查询负载多样化和复杂化的趋势为云服务提供商资源管理带来的挑战,提出一种基于差分进化(DE)和极限学习机(ELM)的方法DE-ELM,对并发查询的性能进行预测。极限学习机用于预测并发查询性能,差分进化算法用于同步优化特征子集和极限学习机结构。该方法仅使用查询编译时信息、无需事先指定特征数目,也无需事先就查询交互的性质、数据库系统的内部运作机制做出先验假设。在合成数据集和真实数据集上进行了详细的实验研究,以评估极限学习机的训练效果、同步优化特征子集和极限学习机结构的效果。结果表明,DE-ELM的平均预测精度高于80%,在一定程度上证明了所提方法的可行性和有效性。

关键词: 并发查询, 性能预测, 极限学习机, 差分进化, 特征选择