FRLDM: A Self-Optimizing Algorithm for Data Migration in Distributed Storage Systems

In distributed storage systems, data migration is an efficient method for improving system resource utility and service capacity, and balancing the load. However, the user accessing is changing over time and the state of a distributed system is in an unpredictable stochastic fluctuation, hence traditional heuristic policy- based methods are hard to work in such environment. This paper proposes a fuzzy reinforcement learning method for online data migration named FRLDM which can enable the systems to self-optimize and dynamically choose the candidate data for migration based on their recent access pattern and the current state of the system, thus minimizing the average access response time. The experimental results prove that FRLDM can improve the accesses performance significantly compared with heuristic policy-based methods.