Buffer Pool-Based DBMS Modeling With Layered Queuing Networks

As the workloads of database management systems(DBMS) becoming more and more diverse and complex, the performance of DBMSs matters. As the backend support system for applications, DBMSs’performance determines the performance of applications. It is an important task to manage performance of DBMSs. A performance model is required to guide DBMSs’performance management usually, therefore it becomes a core mission to create a model of performance for DBMSs.The configuration of buffer pools has a huge impact on the performance of a DBMS. It is impossible to build a performance model of DBMSs without the description of the buffer pools. The buffer pool is a part of main memory, used to temporarily read and update data pages. The data operations of applications are carried out in buffer pools. The reason why buffer pools are necessary in DBMSs is that data access from the main memory is much fast than from the hard disks. Buffer pools can significantly improve the performance of DBMSs by reducing the number of access to hard disks. Therefore, it is necessary to build a performance model for DBMSs with buffer pool description.In this thesis, the method for building performance model is layered queuing network model. Queuing network model describes hardware resources as the service centers, and computer systems are modeled as network of service centers. Layered queuing network model is an extension of queuing network model models both software and hardware resources as the tasks. By analyzing the relationship of the nested call between tasks, the performance bottlenecks can be found. The use of buffer pools involves the query process, the buffer pool management, main memory, hard drives, and other hardware and software resources, which forms a hierarchical relationship. It is difficult to use queuing network model with only single-layer to describe such a system. LQNM is able to describe complex layered systems. LQNM’s nodes can request services to the other nodes as a customer, and provide services for other nodes as a server, while QNM’s node can only provide service for requests. Layered queuing network model is more appealing.Based on the above facts, we propose a DBMS layered queuing network model based on buffer pool description by analyzing of the mechanism of the buffer pool. It is a four-layer model and consists of seven tasks:three software tasks (the query process, the logical read, the physical read) and four hardware tasks (THINK, CPU, the memory, hard disk).Using method of layered the performance model is divided into software and hardware contention model, describing the models’input and output parameters, setting up experimental environment. In the experiments, the target DBMS is DB2i, populated with1G TPC-H benchmark data, and TPC-H queries server as the workloads. Under different buffer pool size, model’s input parameters are acquired by the snapshot monitors of DB2database, and model’s output parameters are acquired by method of layered. The model’s output parameters are the predicted response times. In order to prove the validity of the model, we compare the response time predicted by layered queuing network model and queuing network model, and the actual response time. The experimental results demonstrate that the response time of the predicted value of layered queuing network model is closer to the actual value than the predictive value of queuing network model, which illustrates the rationality and effectiveness of the layered queuing network model.