Large-scale AI clusters impose higher requirements on network scalability, cost, and communication efficiency. The traditional Clos topology suffers from superlinear cost growth when scaling to over 100k GPUs, while the more cost-effective Dragonfly+ introduces "down-up" detours, deadlock risks, and complex routing design. This paper presents Dragonfly-Ultra, a scalable, low-cost network architecture for high-performance AI clusters. Dragonfly-Ultra can scale to over 260k GPUs with only 82% cost and 81% power consumption of a 3-layer Clos architecture. Dragonfly-Ultra optimizes inter-group connectivity to eliminate intra-group detours entirely. Beyond the topological benefits, Dragonfly-Ultra incorporates three key mechanisms to further improve network performance and optimize collective communication, including lightweight dual-waterline adaptive routing for fast congestion mitigation, virtual-link-based deadlock avoidance with lower hardware overhead, and uniform affinity-aware rank placement for balanced inter-group traffic across all phases. Simulation results on a 4k-node cluster show that, compared to Clos, Dragonfly-Ultra achieves up to 18.8% and 39.2% lower completion time for AllReduce and AlltoAll, respectively. Compared to Dragonfly+, the reductions are up to 27.9% and 62.1%, outperforming current mainstream topologies.
Rui Zhuang, Hui Yuan, Junye Zhang et al.· Proceedings of the ACM SIGCO...· 0 citations
Networks with highly dynamic data transmission demands and network topologies are common in real world. A fundamental problem in such networks is achieving scalable traffic allocation to maximize long-term total throughput under link capacity constraints. However, state-of-the-art (SOTA) works lack scalability. This is primarily due to two reasons in large-scale networks: first, they require solving constrained optimization problems online, which leads to high decision latency; second, they rely on reinforcement learning algorithms for policy optimization, which are inefficient in exploration and challenging to train effectively. To address these issues, we propose the Fast Networked Control (FNC) policy framework, which firstly utilizes parallelizable neural network modules to process the state and generate raw decisions, followed by basic operations such as normalizations and comparisons, which do not require iteration or optimization, to obtain decisions that satisfy the constraints. Hence, FNC policy avoids solving constrained optimization problems and supports parallel execution, significantly reducing decision latency. Furthermore, this policy preserves gradient flow and supports backpropagation, which enable us to design an imitation learning algorithm to efficiently train the policy in an end-to-end manner. Experiments in large-scale networks show that our FNC policy achieves an average 8% improvement in demands satisfaction and 10 times reduction in decision latency versus SOTA works.
Zhaoxing Yang, Guiyun Fan, Anjie Cao et al.· IEEE Transactions on Network...· 0 citations