Load Balancing With Proactive Traffic Separation in Data Center Network

Tao Zhang, Yujie Peng, Zhenbo Li, Yuchen Wang, Xidao Luan, Yunsheng Liu, Siyuan Fan, Hui Yin

Abstract: Modern data center when providing services generally induces a vast amount of delay-sensitive and throughput-oriented data flows to be mixed and transmitted in its asymmetric and multi-path enabled physical network, i.e., Data Center Network (DCN). To speed up flow transmission thus facilitating application performance, many data center load balancing schemes are proposed to balance traffic load among DCN parallel paths, expecting to utilize the offered high bisection bandwidth effectively. However, these schemes indulge the heterogeneous data center flows in mixed transmission regardless of their diverse transmission requirements, greatly impeding the improvement of flow-level transmission performance. This paper adopts the idea of Proactive Traffic Separation to design a new data center load balancing scheme, namely PTS, under which the switch output ports are pre-allocated to different types of data center flows by using a time-series forecasting model to learn future traffic arrival intensities. Meanwhile, PTS perceives path conditions in real-time to achieve rational and flexible traffic rerouting. The results of numerous NS2 simulations show that PTS significantly reduces the average and tail flow completion time of delay-sensitive flows, while the throughput-oriented flows’ performance is always maintained at a high level.