Flow Splitter: A Deep Reinforcement Learning-Based Flow Scheduler for Hybrid Optical-Electrical Data Center Network

Abstract

Hybrid optical-electrical switching based data center network (HOE-DCN) has been regarded as a promising architecture for the next generation data center network (DCN). To achieve traffic optimization, the main superiority of HOE-DCN is its capability to offload the long-lived 'elephant' flows by optical interconnections, and transmit the latency-sensitive 'mice' flows by electrical switching. However, most previous works identify and schedule the flows according to a fixed flow size threshold, which can hardly handle the highly dynamic network conditions in recent DCN. In order to achieve more effective flow scheduling in HOE-DCN, in this paper, we propose Flow Splitter (FS), a deep reinforcement learning (DRL) based flow scheduler which enables HOE-DCN to make instant flow scheduling according to the runtime network conditions. To train a more effective DRL agent, we upgrade the DRL method named Deep Deterministic Policy Gradient (DDPG) and propose DDPG-FS, which is capable of learning a high-performance flow scheduling policy in the complex network environment. Through simulation, we prove that our FS can significantly improve the performance of HOE-DCN. Compared with the recent flow scheduling approaches for HOE-DCN, our FS can obviously reduce the average flow complete time of arrival flows, especially the latency-sensitive mice flows.