Research And Implementation Of A Performance Prediction Method For Database System Based On Transfer Learning

With the continuous iterative development of enterprise-level applications,the database provides many functions such as organization,sharing data,and security protection as a key module for data storage and data processing.Database system provides a large number of configuration parameters in order to deal with the task under different scenarios effectively,but due to its large number and complex interaction,database administrators cannot provide optimal configuration parameters for different tasks.And the database version needs to be constantly updated with the continuous changing of system versions and related technologies which reduces the availability of existing,well-learned performance prediction models.Therefore,in the case of version change,the analysis and modeling of database system configuration parameters and their own performance to predict their performance becomes the primary task to optimize the overall performance of the application.At present,there is a certain degree of research in the field of database performance prediction at home and abroad,but existing methods have limited performance prediction to a fixed environment.The problem is that it is not possible to obtain a database system performance prediction model with high accuracy and reasonable cost facing with database component updates and upgrades,the number of samples available and limited modeling time.Therefore,how to use the existing experience and knowledge reasonably to obtain the performance prediction model of the target database system within a limited sample number and time range has become an important issue in optimizing system performance.To solve the above problems,this paper proposes a performance prediction method for database configuration parameters by modeling and analyzing the performance of the database system.This method uses the idea of transfer learning to construct an example of a domain adaptive framework based on adversarial discriminants with reuse of some empirical knowledge in the source domain on the basis of meeting the sample size and time cost constraints,and obtains the performance prediction model of the target database system by combining discriminative feature learning and adversarial learning effectively.More specific introduction of our work is as follows:(1)Give a definition about the database system performance prediction issues.Analyze the factors that affect the performance of the database system,and use mathematical formulas to define the operating environment,workload,database components,and configuration parameters,and perform database system performance modeling.This paper analyzes the problems of database system performance prediction and the shortcomings of the existing solutions,and gives a clear definition of the performance prediction problems in the database version change scenario.(2)Design and implement an adversarial domain adaptive discrimination algorithm based on transfer learning,and build a database system performance prediction model.First step is to do sample initialization according to the configuration parameters,and use stepwise linear regression to perform feature selection.The idea of transfer learning is used to design and implement an adversarial domain adaptation discrimination algorithm that can use the prior knowledge of the source domain database system to generate a target domain database system performance prediction model.An encoder is obtained through adversarial training to encode the target domain database system configuration parameters into the same parameter space as the source domain,so that the performance prediction model in the source domain can predict the performance value of the target domain database system.(3)Design and carry out experimental verification and analyze experimental results.In order to verify the effectiveness,model applicability,and system applicability of the algorithm more comprehensively,this paper proposes three key points in the experimental process: First,in order to verify the effectiveness of the algorithm in the actual working environment,we use six different workload scenarios combined with the YCSB basic test tool to design the experimental content,and try to ensure that the experimental results are analyzed from multiple aspects;implement three different algorithms of random forest,linear regression,and SVM to generate the performance prediction model of the source domain database system and verify the algorithm the model applicability of the algorithm;use My SQL,Cassandra and Redis as the database system to establish performance prediction model,and verify system applicability of the algorithm by experiments.Finally,by comparing two transfer learning methods,it is proved that the algorithm in this paper can obtain a database system performance prediction model with a high accuracy at a lower cost.