ATR: Auto-Tuning Configurations of Redis via Ensemble Learning

Abstract

Redis is one of the most popular in-memory cache databases at present and it has many performance-related configuration parameters. Because of the complexity to explore the configuration space, there is an urgent need of a method that can automatically adjust configuration parameters to quickly and accurately improve the performance of Redis. In this article, we propose ATR: a novel approach to auto-tune the configuration parameters for a given Redis application. The core of ATR is to take Redis configuration parameters as input and to establish an accurate performance prediction model. Then we use the model as a basis to search for configuration parameters with better performance. By systematically comparing various modeling techniques and optimal search algorithms, we employ an ensemble learning algorithm to build the performance model and leverage genetic algorithm to search the optimal configuration parameters. In order to evaluate ATR, we build a Redis cluster with 10 nodes and test the performance by 6 representative applications from YCSB benchmark suite. The experimental results show that compared with the default configuration, ATR increases the throughput by 79% at the maximum, with an average of 48%. ATR reduces the latency by 16.6% at the maximum, with an average of 9.49%.