Data Balancing of Ceph Distributed Storage System Based on Deep Reinforcement Learning

Aiming at the problem of unbalanced data distribution caused by the Controlled Replication Under Scalable Hashing (CRUSH) algorithm used in the Ceph distributed storage system, this paper proposes a data balance algorithm based on deep reinforcement learning for Ceph system. When the traditional method optimizes the uneven distribution of the data of the CRUSH algorithm, it destroys the advantages of the small data movement of the CRUSH algorithm, and the effect is not satisfactory when the number of data copies is large. In this paper, the data balancing problem is transformed into a weight optimization problem of Object Storage Device (OSD) in the Ceph system, and a deep reinforcement learning algorithm based on TD3 (Twin Delayed Deep Deterministic Policy gradient algorithm) is constructed. Experimental results show that the proposed method can improve the balance of data distribution without changing the logic of the CRUSH algorithm and multiple copies.