High-performance datacenter infrastructure is evolving rapidly, yet hardware upgrades alone cannot handle the challenges posed by fluctuating internal traffic patterns. A critical technical pain point in DCN is the frequent congestion caused by transient traffic bursts (Incast), leading to packet loss and excessive queuing delays. Existing protocols are numerous but lack a unified taxonomy, leaving a gap in understanding how to balance throughput and fairness for specific high-concurrency workloads like distributed deep learning. This fragmentation hinders the development of adaptive network management strategies in large-scale cloud environments.
In response to these challenges, the research team from the National University of Defense Technology developed the DCEF framework. This innovation reclassifies congestion control protocols from both temporal and spatial dimensions. The framework identifies three core technical paths: data-driven passive control, credit-based proactive scheduling, and enhanced end-to-end feedback mechanisms. By analyzing these categories, the study reveals how different logic—such as Explicit Congestion Notification (ECN) and Priority Flow Control (PFC)—interact within modern RDMA networks to handle micro-bursts and ensure loss-free transmission.
Research indicates that the DCEF framework provides a clear roadmap for protocol evolution in environments dominated by large-scale AI training. Data analysis suggests that hybrid protocols combining proactive credit management with intelligent feedback are most effective at mitigating Incast congestion and reducing tail latency. This comprehensive survey not only synthesizes a decade of technological progress but also offers a robust technical roadmap for building next-generation intelligent data centers capable of ultra-low latency and autonomous congestion awareness.
DOI:10.1007/s11704-025-40212-y