Predicting database workloads through mining periodic patterns in database audit trails

Abstract

Information about the periodic changes of inten-sity and structure of database workloads plays an important role in performance tuning of functional components of database systems. Discovering the patterns in workload infor-mation, such as audit trails, traces of user applications, and sequences of dynamic performance views, is a complex and time-consuming task. This work investigates a new approach to analysis of information included in the database audit trails. In particular, it describes the transformations of infor-mation included in the audit trails into a format that can be used for discovering the periodic patterns in the fluctuations of database workloads. It presents an algorithm that finds elementary periodic patterns through nested iterations over a four-dimensional space of execution plans of SQL statements and positional parameters of the patterns. It proposes a col-lection of composition rules for the derivations of complex periodic patterns from the elementary and other complex pat-terns and it shows how to use such rules to predict the future workload levels.