Network-Native DNN Architecture Synthesis for Distributed Inference
Fan Yang, Xiaolong Wu, Qi Cao, Fen Hou, Yulin Shao
原文摘要
Distributed deep neural network (DNN) inference typically follows a model-first paradigm, in which a predefined DNN is partitioned and deployed over a physical network. On irregular communication-computation topologies, however, this approach can create transmission bottlenecks and leave distributed resources underutilized. This paper reverses this perspective and asks: given a physical network, what kind of DNN can it naturally support? To answer this question, we propose a network-native DNN architecture synthesis paradigm in which the physical network topology directly defines inter-node neural dependencies. Neural computation is allocated across nodes under communication, computation, storage, synchronization, and latency constraints. We establish structural properties of the synthesis problem and develop linear programming (LP)-based Architecture Synthesis (LPAS), an efficient topology-guided algorithm that produces a feasible architecture with a computable bound on suboptimality. Experiments on physical networks of different sizes show that LPAS achieves relative gaps below 0.6% to the LP-derived upper bound, with substantially lower runtime than direct mixed-integer LP optimization on larger networks. Under matched network deadlines, the network-native architectures activate more compute nodes and utilize a larger proportion of available model-storage capacity than the model-first baseline. After training, the network-native models achieve higher classification accuracy across all evaluated dataset-precision settings, with the average gain of up to 10.69%. This network-native design paradigm thus turns the physical network from a deployment constraint into an architecture design primitive.
cs.IT · cs.DC