Reinforcement Learning Hadoop Map Reduce Parameters Optimization

Abstract

Among the various techniques for enhancing Hadoop performance—such as intermediate data compression, in-memory management, and parameter tuning—dynamic configuration parameter tuning proves to be the most impactful. However, existing approaches face several challenges: limited adaptability to specific application requirements, isolated parameter tuning without considering interdependencies, and inaccurate linear assumptions in complex environments. To address these issues, this study introduces a reinforcement learning-based optimization framework using Q-Learning. The proposed method dynamically adjusts key Hadoop configuration parameters by continuously learning from job execution metrics such as completion time and wait times in map/reduce phases. It employs a reward-based feedback mechanism to minimize the gap between expected and actual performance, ensuring more accurate, adaptive, and holistic optimization. Additionally, the framework integrates a neural network to predict optimal parameter values, further enhancing decision-making. This approach significantly improves execution efficiency and resource utilization, offering robust adaptability across diverse workloads and operational environments, while aligning closely with service level agreements.