Distributed AI model training generates bursty, low-entropy elephant flows that challenge existing single-path transport protocols in multi-stage Clos networks, leading to congestion and inefficiency. Multipath transport emerges as a promising solution, leveraging multiple paths to balance traffic and enhance resilience. However, current multipath RDMA solutions suffer from scalability, congestion control, and load-balancing inefficiencies. This paper introduces Unified Delay-driven Multipath Protocol (UDMP), a novel approach that co-designs congestion control and load balancing using network delay as a unified signal. UDMP employs delay-gradient-based congestion control to precisely resolve unavoidable congestion. Moreover, UDMP leverages delay-assisted load balancing to shift traffic across paths with minimal latency adaptively, maintaining throughput when encountering avoidable congestion. A novel Token Pool design integrates these components, eliminating per-path state overhead while achieving fine-grained traffic distribution. Implementations on DPDK and NS3 demonstrate that UDMP achieves up to 2x higher throughput and reduces flow completion times by up to 30% compared to state-of-the-art methods like MPRDMA and QP-Scaling. These results highlight UDMP’s effectiveness in meeting the stringent performance requirements of modern distributed AI training workloads.

Cloud providers require high-throughput traffic to test the effectiveness of congestion control (CC) configurations (i.e., CC algorithm selection and their parameter settings) in networks. A network tester capable of evaluating CC configurations needs to fulfill the following requirements: (R1) Capable of generating traffic with CC behaviors. (R2) Ability to customize CC algorithms. (R3) High throughput CC traffic generation. However, existing network testers fail to meet these requirements simultaneously. The paper presents Marlin, a novel high-throughput network tester designed for CC evaluation. Marlin leverages a high-throughput, low-programmability device to amplify the traffic generated by a low-throughput, high-programmability device. The low-throughput device is responsible for complex computational tasks, such as running CC and flow scheduling algorithms, and communicates with the high-throughput device at a high frequency using small packets to instruct it to generate high-throughput traffic with CC behaviors. This hybrid approach allows for customizable, high-throughput CC testing. Our experiments demonstrate that Marlin can accurately emulate CC behaviors and replicate real-world scenarios. Marlin can generate 1.2 Tbps of CC traffic using a single programmable switch pipeline and one 100 Gbps port of an FPGA NIC, supporting up to 65,536 concurrent flows.

State-of-the-art inter-datacenter WANs rely on centralized traffic engineering (TE) to improve the network performance, where TE computation is a periodical procedure and timely performs routing configurations (i.e., enforces TE polices via add, remove and modify forwarding rules) in response to the changing network conditions. The TE computation determines the routing configurations corresponding to the current network conditions and the network update operations change the routing configurations from last TE to current TE solution. Existing works take centralized TE computation and network update as two individual optimization procedures, which inevitably leads to suboptimal solution in the long run. In this paper we initiate the study of online joint optimization on TE computation and network update with the objective of minimizing the sum of TE cost and network update cost. We formulate this problem as an optimization program and propose a set of provable online algorithms with rigorous competitive and regret analysis. Trace-driven simulations on two empirical topologies demonstrate that our algorithms can significantly decrease the total cost.