Research On Workload Adaptation Technology For DBMS

Workload adaptation of database systems is initiated for optimizing the performance of large database systems. Much attention has been given to the optimization of the internal software and hardware resources of the DBMS itself in the literature, whereas less research was conducted on the workload perspective. In 2006, the researchers from Queen's University of Canada and IBM proposed the idea of workload adaptation of DBMS which achieves the performance optimization of the DBMS by strategically choosing the optimized scheme of timing and sequence of the execution of the workload requests with constant resources consumption of the application system. Workload adaptation of DBMS belongs to the research scope of autonomic computing of database, which is one of the most important research subjects today.The workload of database is instantaneous, complex and variable. Database workload control, real-time classification and performance prediction are the core research contents to implement workload adaptation.The systematic study on the algorithms of workload feature classification, database performance prediction, workload control and the database workload adaptation functional framework etc are summarized in this dissertation. The key content and innovation are as follows.In the workload classification process, the cluster based on feature vectors (CFV) algorithm is proposed. This algorithm uses cluster method to extract the relationship of load's dynamic and static characteristics, executes load classification dynamically, which effectively avoid the problems associated with the static and experiential load classification. CFV reduces the workload feature vector dimensions, and reduces the complexity of matching. A refined Incremental Cluster based on Feature Vectors (ICFV) algorithm is also initiated based on the CFV algorithm. ICFV reduces time complexity of classification greatly and improves the efficiency of operation by using online incremental cluster method to classify loads. Synthetic control chart time series data in UCI is used to do some experiments using MATLAB, and the results show that the algorithm supports the performance requirements of the workload-online-classification.In the process of the performance prediction of the DBMS, a layered queuing network model based Rendezvous Mean Value Analysis (RMVA) is initiated. RMVA introduces the concept of rendezvous and solving the layered queuing network model by using approximate average value analysis method. It solves the problem that prediction model can not describe the complex relationships within the database accurately, and it also reduces the space and time complexity of solving the performance model. The experimental results show that the RMVA method yields much faster calculations on the average performance parameters for the DBMS, and it will not cause the explosion of the status space coupled with the increasing numbers of workload.In the process of workload control, a refined NSGA-II algorithm which is suitable for workload control of DBMS is proposed. This algorithm modifies the invalid individual strategy from the strategy population of workload control, and optimizes the valid strategies, making each of the individual strategy to meet the workload criteria and to be approximately optimized within the limit of performance constraints. It also meets the demand of improving the quality of resolution and speeding up the iteration convergence. The test in MATLAB using the standard data set shows that, comparing with NSGA-â…¡algorithm and the case without load control, refined NSGA-II algorithm not only improves the quality of resolution, but also speeds up the iteration convergence.To finalize the workload adaptation architecture, we introduce the knowledge base into the framework, and set up the workload adaptation architecture for DBMS(WAAD) and recognizing the knowledge base as the core component. The architecture includes knowledge base, workload classification component, performance prediction component, workload control component, and the system-monitoring component etc. The results of the functioning of each component are stored in the knowledge base as records of knowledge, which carries out necessary deduction and analyzing of the recorded knowledge. The introduction of the knowledge base complements the workload adaptation architecture, solves the problem that the adaptation of previous framework is inadequate, and reduces the computing scale and complexity of workload classification algorithm and workload control algorithm owing to the support of knowledge base.According to the workload performance optimization requirements of the power marketing management DBMS, this dissertation uses the above-mentioned CFV, ICFV, RMVA and refined NSGA-II Algorithms and developed workload adaptation prototype system in the power marketing DBMS. After experimental testing, the system reaches the workload performance optimization targets of the power marketing management DBMS, which approves the effectiveness of the algorithms proposed. It is believed the research progressing made in this dissertation will also provide reference for further researches on the workload adaptation technologies.