Intelligent Transaction Scheduling Based On Reinforcement Learning

With the development of database technology and computer hardware,current hardware platform for database running has gradually been transferred to a platform composed of multi-core processors and large-capacity low-latency storage,which resulting in in-memory database.In-memory databases typically schedule transactions randomly between different threads to maximize concurrence,without considering the potential conflicts of multiple transactions executing in parallel.However,an important reason for the degradation of the OLTP performance of the in-memory database is that transactions are aborted due to collisions when access to the same data.If a method can be adopted to reduce the probability of transaction collision,the system performance can be improved.The rapid development of machine learning provides opportunities for autonomous databases.Transaction collisions are difficult to predict,and the conditions for collisions are very complex.Therefore,machine learning methods can be used to learn transaction scheduling tasks.Training machine learning models through historical experience data,and then using model output to better schedule transactions can improve DBMS performance.In this article,an intelligent method of in-memory database transaction scheduling,combined with reinforcement learning,was explored.Before training and using the reinforcement learning model,first perform transaction feature extraction as input data.The existing methods only focused on feature extraction for equivalent queries,which lack of versatility.This paper proposed a feature extraction method that can simultaneously perform feature extraction on equivalent query and range query transactions.And corresponding feature description method,feature extraction algorithms and feature standardization method were proposed.Next,a DQN model suitable for transaction scheduling was constructed.First,the overall structure of the reinforcement learning model was introduced.Then,the limitations of Q-Learning were proposed,and a DQN algorithm that was more suitable for transaction scheduling task was derived.The original DQN algorithm needed to be adjusted for transaction scheduling tasks.This article described in detail the key details of the DQN algorithm and adjustments made to the transaction scheduling problem.Then the key details of the DQN algorithm and the adjustments made to the transaction scheduling task were described in detail.The output results of different machine learning models had different representations.Based on the DQN model,this paper designed a basic scheduling algorithm.According to the-greedy strategy,a scheduling queue was randomly selected,or the optimal queue was selected by using the calculation results of the DQN model.In order to further optimize the scheduling algorithm,a blocking strategy was designed to prevent the average response time of a queue from being too long due to too many current transactions in this queue.Finally,the intelligent transaction scheduling method based on reinforcement learning was implemented in the Noise Page database system,and the effectiveness of this method was verified through experiments by running two benchmarks.