Research On Database Parameter Tuning Method Based On Load Prediction

Database as the infrastructure of the information age,was an important support for every organization,institution,application and business.The parameter configuration in the database management system was an important factor affecting the operation as well as the performance of the database.These parameters determined whether stable services and overall system performance could be provided under different business scenarios,and the optimized parameter configuration could greatly improve the availability of the database and the application system compared with the default configuration.At present,most of the database parameter tuning was based on the current workload status,but the workload changed more frequently,and the current configuration parameters could not be well adapted to the workload demand in the future period,resulting in low performance of the database or system crashes and other problems.Therefore,database parameter tuning needed to take into account the trend of load change,first predicted the load change,and then performed database parameter tuning based on the load change.In this thesis,we focused on the study of deep reinforcement learning parameter tuning model under load prediction,aiming to reduce the problems of outliers and deep learning load prediction model out missing medium and long sequence information,not focusing on global information dependence,and the problems of large computational error and long tuning time due to the parameter tuning model based on deep reinforcement learning method due to its inherent maximization of output action and estimation,and constructed a load prediction-based database parameter tuning framework,which mainly included the following two research points.(1)To address the problem that the unidirectional long and short-term memory network in load prediction only utilizes positive information and medium and long sequences of lost information,a load prediction method based on a bidirectional long and short-term memory neural network combined with an attention mechanism that filters outliers is proposed.The collected database workload metrics descriptions were analyzed by using database workload historical data samples for prediction.First,the presence of outliers in the load could greatly affect the prediction results,and the isolated forest anomaly detection method was used to filter them.Secondly,the reverse information was also beneficial to improve the accuracy of the model for sequence prediction.Finally,the attention mechanism was combined to ensure attention to global information,and the dependence of important information in the sequence was used to predict future database workload changes.(2)To address the problems of large computational errors and slow convergence in the deep deterministic strategy parameter tuning model in predictive load tuning,a load prediction parameter tuning method based on a double-delay deep deterministic strategy is proposed.Firstly,four Critic networks were used to update alternately,and the minimum estimated values calculated by two Critic target networks were selected to eliminate the overestimation problem,and then delayed update of Actor network was used to make the Critic network more deterministic before Actor acts to optimize the search of parameter configuration.Finally,the database parameter tuning effect was improved with accelerated model convergence time.Finally,in the method implementation stage,the load prediction method,load prediction parameter tuning method and corresponding techniques proposed in this thesis were used to implement FTune,a database parameter tuning framework based on load prediction.Collected database parameters and performance information and performed factor analysis to reduce and cluster to remove redundant terms using LASSO regression method to generate a list of importance parameters,and tuning according to the importance degree of parameters to reduce the cost of searching the network space.The experimental results showed that the throughput performance of this thesis has 22.5% improvement and 36.7% latency reduction compared with the original load tuning.The parameter tuning method based on load prediction in this thesis completed the accelerated iteration in the parameter tuning process compared with the parameter tuning method based on current load,and responded to the load change in advance.