MP-CreditINT: Enhancing multi-path RDMA transport with credit-based congestion control and in-band network telemetry

Remote Direct Memory Access (RDMA) has played a critical role in recent Large-Language-Model (LLM) training workloads by enabling low-latency, high-throughput communication across GPUs. To further improve network efficiency, multi-path RDMA transport has received increasing attention. Given the limited on-chip resources of RDMA NICs, MP-RDMA stands out as a state-of-the-art multi-path transport by adopting memory-efficient mechanisms for congestion and out-of-order control. However, MP-RDMA relies on ECN, which provides only coarse-grained binary congestion signals, resulting in limited congestion control capabilities. Our experimental observations reveal intrinsic limitations of MP-RDMA, such as slow bandwidth probing, large oscillations in queue buildup, and transient congestion, etc. These limitations reduce the network efficiency, especially under All-To-All communication patterns that are becoming increasingly dominant with the evolution of Mixture-of-Experts (MoE) models and Expert Parallelism. To address these limitations and retain the memory-efficient property of MP-RDMA, we propose the MP-CreditINT. This approach re-architects MP-RDMA using credit-based congestion control and in-band network telemetry (INT). We systematically enumerate and address the architectural and algorithmic challenges arising from this transformation, including explicit path control and path symmetry, INT-based data-clocking credit window control, and robustness against feedback loop breakage. Then, we evaluate MP-CreditINT using micro-benchmarks, heavy incast traffic, and LLM training workloads. Simulation results demonstrate that when compared to MP-RDMA, MP-CreditINT achieves 6–38× faster ramp-up speed and 8–16× faster fairness convergence, while maintaining near-zero out-of-order degree. In heavy incast scenarios, it achieves superior fairness and near-zero queue buildup, whereas MP-RDMA exhibits exponential queue growth with increasing incast scale. Finally, under two representative LLM training workloads, All-Reduce and All-To-All, MP-CreditINT reduces completion time by 5%–8% and 7%–13% respectively when compared to MP-RDMA, demonstrating its benefits in LLM training. • We conduct an in-depth experimental analysis of MP-RDMA and identify several fundamental limitations rooted in its ECN-based design, including slow ramp-up speed, large out-of-order degree, and the inability to effectively handle incast congestion. • We propose the overall architecture of MP-CreditINT, which retains MP-RDMA’s hardware-efficient philosophy while fundamentally re-architecting its control loop from Data-ACK to Credit-Data paradigm. We systematically enumerate and address the architectural and algorithmic challenges arising from this transformation, including explicit path control and path symmetry, INT-based data-clocking credit window control, and robustness against feedback loop breakage. • Experimental results show that MP-CreditINT achieves 6–38× faster ramp-up speed and 8–16× faster fairness convergence than MP-RDMA, while maintaining near-zero out-of-order degree. In heavy incast scenarios, MP-CreditINT achieves superior fairness and near-zero queue buildup, whereas MP-RDMA exhibits exponential queue growth with increasing incast scale. Under two representative LLM training workloads, All-Reduce and All-To-All, MP-CreditINT reduces completion time by 5%–8% and 7%–13% respectively when compared to MP-RDMA.