Existing classifications slice the congestion control landscape along single dimensions. The temporal view separates reactive protocols, which respond to congestion after it manifests through packet loss or increasing latency, from proactive protocols that adjust rates based on predicted congestion before queues build. The arbitration-subject view distinguishes pure end-to-end schemes, where edge hosts alone make decisions, from switch-assisted or network-arbitrated mechanisms that offload intelligence into the fabric. While useful, these one-dimensional labels obscure the nuanced communication patterns that truly differentiate protocol families. The HiNA team argues that the heart of any congestion control algorithm lies in how congestion signals are generated, propagated, and acted upon, and whether the control loop relies on a separation between data and credit transmission.

The DCEF framework begins with a critical architectural insight: in many modern protocols, the flow of data packets and the flow of control credits need not travel the same path or follow the same timing. A credit represents the receiver’s or network’s permission to send a certain amount of data, and decoupling its transmission from the data itself opens up a rich design space. DCEF categorizes congestion control along four axes derived from this separation: the nature of the congestion signal (explicit versus implicit), the path of the signal (in-band alongside data or out-of-band on a separate channel), the entity that generates the signal (receiver, switch, or a combination), and the granularity of the feedback (per-packet, per-flow, or aggregate). By mapping protocols onto these dimensions, the framework illuminates why certain designs excel in particular environments while others fail to converge under incast patterns typical of distributed storage workloads.

The team applied DCEF to a representative set of classic protocols spanning the years 2010 to 2022, including landmarks such as DCTCP, TIMELY, SWIFT, and HPCC. Their analysis reveals that reactive protocols using explicit congestion notifications from switches, like DCTCP, place the signal generation responsibility entirely on the network fabric and carry that signal in-band via ECN marks, whereas credit-based proactive schemes like ExpressPass issue credits from receivers off-band, thus preventing queue buildup altogether. The survey not only classifies these protocols but also provides a comparative evaluation of their performance in terms of throughput and tail latency, their convergence speed under dynamic traffic, and their deployability constraints such as required switch hardware modifications. The DCEF mapping shows that while switch-generated explicit signals offer fast convergence, they often demand custom ASIC support that hampers incremental deployment.

A deeper technical dimension explored by DCEF is the tension between convergence stability and signal granularity. Protocols that rely on per-packet credit return, for instance, achieve fine-grained rate matching but may become unstable if credit loops or deadlocks are not carefully avoided. On the other hand, flow-level feedback schemes reduce signalling overhead but are ill-suited to microbursts. The framework contextualises these trade-offs, making it a valuable tool not just for taxonomy but for guiding future protocol design. The survey also touches on the interplay between congestion control and emerging network interface cards, specifically SmartNICs, which can perform on-path credit accounting and signal generation at line rate, effectively blurring the traditional boundary between end-host and switch responsibilities.

Looking ahead, the authors foresee that the DCEF framework will become especially relevant as SmartNIC adoption accelerates. With programmable data planes and hardware-accelerated credit management, the distinction between in-band and out-of-band signal transmission may dissolve, enabling hybrid protocols that adapt their congestion signal pathways on the fly based on network conditions. The team envisions congestion control architectures where the credit debt of each flow is tracked by the NIC, while the switch provides only lightweight congestion state hints, a design that neatly separates mechanism from policy. Such advancements could finally shrink the buffer size requirement without sacrificing throughput, tackling the root cause of the buffer-bandwidth divergence that ignited this line of research in the first place.

The survey’s publication marks a timely conceptual consolidation in a field often driven by incremental tweaks to existing algorithms. By establishing a clear language for describing how data and credit flows interact, DCEF may accelerate the development of next-generation congestion control protocols that are both high-performing and pragmatically deployable. As the HiNA team concludes, the classification framework not only structures past work but also defines a forward-looking blueprint for the co-design of hardware and software in the data center, a blueprint that could be pivotal for sustaining the exponential growth in network bandwidth without hitting the buffer wall.

Subject of Research: Not applicable
Article Title: End-to-end congestion control in datacenter networks: a survey
News Publication Date: May 15, 2026
Web References: https://doi.org/10.1007/s11704-025-40212-y
References: Frontiers of Computer Science, DOI: 10.1007/s11704-025-40212-y
Image Credits: HIGHER EDUCATION PRESS
Keywords: congestion control, data center networks, DCEF framework, SmartNIC, credit-based flow control, buffer sizing, network architecture