Joint Computing and Network Resource Optimization for Distributed Training Using PPO

ObjectiveWith the rapid growth of large-scale model parameters, a single data center is increasingly constrained by computing capacity, power supply, physical space, and network bandwidth. Interconnection among intelligent computing centers has therefore become essential for cross-domain parallel training. However, in cross–data center scenarios, heterogeneous communication flows generated by data, tensor, and pipeline parallelism are closely coupled with routing, spectrum allocation, and link occupation in the underlying optical network, which may increase communication latency, aggravate local congestion, and lead to imbalanced computing resource utilization.MethodsTo address these issues, this paper develops a joint computing-network optimization model for distributed large-model training across interconnected intelligent computing centers. It analyzes the communication mechanisms of data, tensor, and pipeline parallelism, and builds a computation–communication time model covering stage computation cost, inter-layer communication cost, and data-parallel global synchronization cost. Model stage partitioning, node selection, computing and storage constraints, together with optical-network routing and spectrum allocation constraints, are then formulated as a unified combinatorial optimization problem. The scheduling process is further modeled as a Markov decision process, and a Proximal Policy Optimization (PPO)-based intelligent scheduling algorithm is proposed. Within this framework, an Actor-Critic structure generates stage partitioning and node-placement decisions, while feasibility checking and latency feedback are incorporated to support iterative policy refinement.ResultsSimulation results show that, compared with static, nearest-node, resource-aware, and greedy strategies, the proposed PPO-based method can better match training-task requirements with the underlying optical-network state under multi-task and high-load conditions. It reduces request blocking probability and average training iteration latency while maintaining high computing-resource utilization and low variance of node computing load.ConclusionThe proposed method realizes coordinated optimization between distributed large-model training tasks and optical-network resources, providing a feasible solution for efficient training scheduling and computing-network resource management in cross–data center intelligent computing interconnection scenarios.