Mjolnir: A framework agnostic auto-tuning system with deep reinforcement learning

Choosing the right setting for big data frameworks is an important yet difficult task. These frameworks come with a complex set of parameters that need to be tuned to achieve the best performance in terms of throughput and latency. Learning-based auto-tuning methods using traditional machine learning models might not be effective for the task because they require huge amounts of high-quality training data, which is time-consuming and very expensive. A good alternative would be to consider reinforcement learning methods to train an intelligent agent through trial and error. In this context, we propose a framework-agnostic auto-tuning system implementing an actor-critic algorithm namely TD3 (Twin Delayed Deep Deterministic Policy Gradient). We show that the agent can find an optimal configuration in a continuous high-dimensional search space with a limited number of steps. We conducted extensive experiments on Apache Spark, under different workloads from the HiBench, TPC-DS and TPC-H benchmarking tools. In this paper, we give a detailed representation of the reinforcement learning environment and show the best design through experiments. Results showed that our approach outperforms the state-of-the-art tuning methods and can improve the performance of spark workloads over the default configurations by up to similar to 77% with an average of similar to 45%. It also showed a promising adaptation behaviour to workload variation during evaluation.