Optimizing Video Quality in Distributed Storage Systems via Deep Reinforcement Learning (DRL)-Based Adaptive Replication

초록

Distributed storage systems, such as Hadoop distributed file system (HDFS), are widely used for video storage due to their outstanding scalability. However, they frequently face challenges related to data unavailability arising from issues like network disconnection, server downtimes, and storage failures. This necessitates file replication, leading to significant storage requirements. To address this, we propose a novel deep reinforcement learning (DRL) algorithm that relies on the tradeoff between video quality and storage demands. We first formulate an optimization problem with the objective of maximizing video quality while constraining storage requirements necessary for replication. The video storage system is then modeled with time-varying video streaming workloads as the DRL environment, where the agent determines the placement of replica files without foreknowledge of future storage availability and video popularity. To address this uncertainty, we use a deep double-Q network (D3QN), which includes an action space that finds the number of replicas for each file, an observation space featuring storage utilization and file placement, and a reward model calculating the expected video quality under various data unavailability probabilities. The implementation of our method is examined within the HDFS. Experimental results show that our method improves video quality by up to 39% compared to benchmarks, achieves quality comparable to Oracle when all bitrates are accessible, and even surpasses HDFS's triple redundancy method while using only 20% of the storage space. © 1991-2012 IEEE.

키워드