| While hardware and software costs drop dramatically, DBA cost and the complexity of DBMSs increase, the total cost of ownership (TCO) of DBMSs is increasingly dominated by people costs. The only solution to reduce such costs is self-managed database that is capable of managing and tuning itself. When tuning the DBMSs, the type of the workload presented to it is a key consideration, because the strategies for resource allocation are different depending on whether it is Online Transactional Processing (OLTP) or Online Analytical Processing (OLAP). A DBMS typically experiences changes in the type of workload during its normal processing cycle. Ideally, the database system shall be able to automatically adjust to such changes. To achieve it, automatic recognition of database workload should be studied firstly.In this thesis, twelve workload characteristics which can distinguish between OLTP and OLAP are selected. The train data set is obtained through TPC-C and TPC-H benchmark test which simulate the two different types of workload respectively. Then a classification model is built based on C4.5 and Boosting algorithm to automatically recognize the type of workload. The algorithms are improved to implement that the classification result is the membership probability of every class. The results of three experiments show that this classification model meets the requirements of automatic recognition of workload on prediction accuracy, robustness and genericness. They also show that the way to implement the classification result is the membership probability of every class is feasible.A workload prediction architecture is proposed for the system which workload shows the cyclical variation. The architecture has three parts: training data model, offline module and online module. It helps the database forecast when the workload may change and the changing trend and then verifies it at the corresponding time using the classification model to decide whether to tune its resources accordingly. |
PDF Full Text Request