Flow-Level Traffic Scheduling in Data Center Networks

Based on software-defined networking (SDN), many online service providers deploy novel online traffic engineering (TE) schemes in their data center networks. The online TE epoch usually ranges from several hundred milliseconds to a few minutes, which is shorter than the duration of some flows. These flows may suffer inconsistent quality-of-service (QoS) between TE epochs, which leads to degradation in flow utility functions and the total network utility. Although effective in packet-level QoS improvement, conventional TCP variants and queueing designs cannot address the QoS inconsistency since flow-level information, including flow service history and flow utility function, is not considered. This paper proposes a flow-level traffic scheduling (FLTS) method to improve flow QoS consistency and network utility. We propose a flow measurement mechanism that factors the service history and residual demand into the flow priorities. For background flows, we establish a G/G/1 queueing model in traffic sources and propose a traffic rate control mechanism to minimize the queueing time. Simulation results demonstrate that FLTS outperforms two typical conventional approaches with an 18% to 29% improvement in network utility under heavy traffic loads.