Multi-task sequence learning for performance prediction and KPI mining in database management system

摘要

Predicting future performance curve and mining the top-K influential KPIs are two important tasks for Database Management System (DBMS) operations. In this paper, we propose a multi-task sequence learning approach to address the two tasks in a uniform framework. The proposed approach adopts a Long Short-Term Memory (LSTM) based deep neural network model that uses multilevel discrete wavelets transform and LSTM-based Seq2Seq forecaster to capture the features in both time and frequency domains from high dimensional time series, and achieves multi-step performance prediction and top-K KPI mining concurrently. The performance of the proposed multi-task sequence learning approach is evaluated based on two real-world DBMS datasets, which shows that the proposed approach achieves the lowest mean absolute error and root mean squared error in predicting performance scores, and significantly outperforms the state-of-the-art algorithms in both learning tasks.