With the growing demand for big data storage and processing, distributed storage systems with heterogeneous devices have become the majority in cloud data centers. However, hardware heterogeneity and workload variety make it challenging to maintain optimal performance in those storage systems. In this work, we present a learning-based control method that optimizes the performance of a distributed storage system. Specifically, to provide automatic parameter tuning upon dynamic workload patterns on a tiered storage architecture, we employ deep reinforcement learning (RL) and implement a simulation environment for a Ceph storage system. Through simulation tests, we demonstrate our RL method shows better performance than other heuristic approaches for a task of load balancing based on the primary affinity settings in Ceph.