Research On Automatic Configuration Tuning For Multi-model Database OrientDB

In recent years,multi-model database is a new research direction in the field of database.Orient DB,as an emerging multi-model database,the default configuration of which cannot achieve the best performance,so it is often necessary to tune the configuration for the actual application workload.More importantly,the existing tuning experience on the single model database cannot be directly migrated to Orient DB.It is urgent to study the automatic configuration tuning technology for Orient DB to reduce the difficulty of configuration tuning,at the same time improve Orient DB performance and make full use of it.This paper investigates automatic configuration tuning for Orient DB,selects eleven configurations from six aspects: memory setting,connection pool setting,timeout setting,parallel query setting,write-ahead logging setting,and transaction log setting to tune.This paper models the automatic tuning process of Orient DB configuration based on deep reinforcement learning,designs and implements a framework named OCTune.OCTune builds a configuration tuning model based on deep deterministic policy gradient.In theory,the modules involved in Orient DB configuration tuning are mapped one by one with the six key elements in deep reinforcement learning,and an effective reward function is designed for the application scenarios of Orient DB to guide the training of the tuning model;in the implementation,a stress test tool named MMBench is designed to generate the workload of cross-data model queries,which enriches the performance test scenarios of Orient DB,a state monitor named OSMonitor is designed to monitor and process the running state of Orient DB in real time,and a performance statistics tool named OPIndicator is used to evaluate the performance changes of Orient DB.OCTune uses MMBench,OSMonitor and OPIndicator to collect the training data,achieving the training of the tuning model without the historical tuning experience of human.Finally,the tuning effect and adaptability of OCTune are tested under four workloads of Orient DB,and considerable results are obtained.Experimental results show that,under workloads of CRUD and GINSERT,after being tuned by OCTune,the performance of Orient DB could be improved by 3.6%～54.9%.