A traffic control mechanism for concurrent flows in data center networks

Given the ever-increasing demand for data transmission, congestion in data center networks has become an increasingly critical issue. Existing solutions only adjust the congestion window and employ ECN feedback at the transport and network layers, applying the same forwarding strategy to all types of flows. When a large volume of highly concurrent flows is simultaneously injected into a bottleneck link, the shallow buffers within switches become exhausted, resulting in TCP timeouts or packet loss, which in turn leads to TCP incast. To address this challenge, we propose an innovative application-layer traffic control mechanism termed DCNFP. This mechanism manages connection requests in batches by integrating proactive connection management with ECN marking and congestion window adjustments. DCNFP uses priority-based multi-queue management techniques to centralize micro-burst flow tolerance method into short flow queue management. As a link nears congestion, DCNFP terminates connections with a high likelihood of timeout, thereby achieving precise regulation of link traffic and alleviating the load on bottleneck switches. Simulation results on the NS3 platform demonstrate that, compared with DCTCP, DCNFP significantly decreases packet loss and improves overall network throughput.