Long and short flow buffer management in data center networks

Abstract

With the advent of multi-source interactive technologies, data center networks are generating a substantial number of short flows that necessitate rapid feedback. However, short flows and long flows have inherent differences. When they share the same buffer space, the performance of short flows is adversely affected by long flows. This results in delayed feedback for short flows, which in turn degrades the user experience. Currently, both short and long flows are managed through dynamic threshold settings or scheduling. However, the influence of long flows on short ones persists. This paper introduces a novel buffer management algorithm Long and Short Flow Buffer Management (LSBM) by dynamically fine-tuning the PFC threshold. LSBM assesses flow congestion by analyzing both space occupancy rates and flow rates while distinguishing threshold controls based on the distinct characteristics of long and short flows to facilitate optimal allocation of available buffer space. In large-scale simulations based on NS-3, experimental results indicate that compared to the original Data Center Quantized Congestion Notification (DCQCN), DCQCN+LSBM achieves lower completion times for short flows without compromising throughput. Furthermore, relative to Dynamic Threshold (DT) policy and Active Buffer Management (ABM), LSBM demonstrates improvements in average finish time for short flows by 52.7% and 26.5%, respectively, thereby achieving dual objectives in flow control and congestion management.