A predictive replication for multi-tenant databases using deep learning

摘要

The service level agreement (SLA) is an agreement between clients and the service provider, which defines the minimum performance and availability requirements for software. The service provider should have effective strategies for placement, migration, and replication of the tenants to reduce their operational costs, maximize the utilization of their hardware and software resources and accordingly meet the SLA requirements with slight SLA violations. In this research, a clustered-based multi-tenant database management system (CB-MT DBMS) is proposed. Additionally, a dynamic proactive provisioning technique is built using three different prediction models: The Recursive Window Forecasting Autoregressive Integrated Moving Average (ARIMA) model, Exponential Moving Average (EMA) model, and the proposed Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) cells model. Various experimental scenarios with different datasets have been conducted to prove that the RNN model with LSTM cells is a promising solution in multi-tenant environments, where the tenants have irregular workload patterns. Experimental results firstly show that the RNN model accuracy is superior to their counterparts (i.e., ARIMA and EMA models) when applied to multi-tenant database workloads generated using TPC benchmarks, as it reduces the prediction error value which is computed using the Mean Absolute Percentage Error (MAPE) and the Root Mean Square Error (RMSE) metrics. Secondly, Experimental results prove that the RNN prediction model accuracy is superior to their counterparts for detecting SLA violation values and windows using different SLA values.