In data center networks, the growth rate of switch buffer size is much smaller than that of network bandwidth capacity. This results in frequent network congestion, which requires congestion control. Currently, the classification of end-to-end congestion control protocols is based on single dimensions such as time or congestion arbitration subjects. From a temporal perspective, end-to-end congestion control protocols are divided into reactive congestion control protocols and proactive congestion control protocols. From the perspective of congestion arbitration subjects, it is divided into end-to-end congestion control, switch congestion control, and network arbitration. However, this classification method is relatively coarse-grained for classifying congestion control protocols. A research team proposes a new classification method for congestion control protocols.
To better understand and compare existing congestion control protocols, a research team named High performance Network and Architecture (HiNA) led by Dezun Dong published their new research on 15 May 2026 in Frontiers of Computer Science, which is co-published by Higher Education Press and Springer Nature.
Based on the above classification method, this team proposes a new congestion control architecture that distinguishes between data and credit transmission, and summarizes four different classification methods according to congestion signal classification, called DCEF framework. The proposed DCEF framework delineates some classic congestion control protocols spanning from 2010 to 2022 and classifies them in accordance with the proposed classification method. Additionally, the work also conducts a comparison of the performance, convergence and deployability of these congestion control protocols.
The work delves into the future development trend of the DCEF framework. Particularly, with the widespread utilization of SmartNIC, the functions supported by switches have grown more intricate. This provides valuable insights for the development of congestion control protocols in the future.

DOI:10.1007/s11704-025-40212-y