Dynamic Control of Storage Bandwidth Using Double Deep Recurrent Q-Network

摘要

We propose a novel approach to optimize the performance of a large scale physical system by mapping the performance optimization problem into a reinforcement learning framework. A reasonably efficient manual bandwidth control for large storage servers seems to be a difficult task for system administrators, but a dynamic bandwidth control can be effectively learned by a reinforcement learning agent. We adopt a combination of Double Deep Q-Network and a Recurrent Neural Network as our function approximator to identify the extent of bandwidth control (actions) given the state representation of a storage server. Allowing the agent to control the amount of allowable bandwidth to each logical unit within a filer has shown to enhance throughput as-well-as reduce the overload duration of storage servers.