Robust Time-Shared Backup Strategy for Training of Large Models in Cloud-Edge Optical Networks

With the rapid advancement of artificial intelligence (AI), the scale of neural network models has grown dramatically, and the computational resources required to train large-scale models (e.g., GPT-4, PaLM) far exceed the capacity of a single device. Cloud-edge optical networks leverage the collaborative advantages of cloud and edge servers to enhance resource utilization and improve real-time task performance. However, this approach heavily relies on link stability for parameter synchronization and gradient transmission. Any interruption in connectivity can disrupt training tasks and increase latency, potentially causing training failure. To address this challenge, this paper proposes a time-shared backup path protection (TSBPP) strategy to optimize link resource utilization in cloud-edge optical networks. TSBPP consolidates parallel model training time by reserving protective resources during idle periods. By facilitating resource sharing across the time dimension, TSBPP effectively reduces reliance on backup resources. In instances of link failure, TSBPP enables swift restoration of communication paths, thereby ensuring continuity in distributed training tasks. Numerical results demonstrate that compared to traditional 1+1 backup, TSBPP achieves significant advantages in terms of resource utilization and fault recovery efficiency.