Automatic knobs-tuning for DB2 using deep reinforcement learning

Abstract

Modern database management systems have hundreds of different configuration parameters (knobs) that control various aspects of how they behave and perform. These knobs must be properly tuned in order to maximize the performance of the database for a given query workload. Traditionally, database administrators would be responsible for database performance tuning. However, manual configuration tuning is a difficult process for humans, as there are hundreds of different inter-dependent knobs to be tuned. Different queries and workloads also benefit from configurations differently, there is no one single database configuration that can fit all scenarios. We propose BLUTune, a system to automatically produce effective knob configuration for IBM DB2. BLUTune utilizes deep reinforcement learning and features a unique transfer-learning approach to training which allows for fast learning. In experimental validation, BLUTune demonstrates its capability of producing effective configurations across differing sizes of the TPC-DS OLAP benchmark in a timely manner.