To pursue better predictive performance in deep neural networks (DNNs), the size of learning models tends to increase, resulting in high computational requirements and training latency. In this paper, we investigate the problem of device placement to accelerate large-scale distributed DNN training. To address the challenge of finding an effective scheme due to its NP-hardness and the ever-increasing model size, we propose novel operator fusion and co-location schemes to reduce the search space while minimizing subsequent performance loss in training latency, enabling efficient device placement. We evaluate the performance of our design with real-world DNN benchmarks, and the results show that, compared to state-of-the-art approaches, our design achieves up to a 35 % reduction in DNN training latency and an order-of-magnitude improvement in placement search latency.

In this paper, we propose a mobility-adaptive digital twin (MADT) framework for network traffic prediction in a post-disaster scenario where some affected terrestrial base stations (BSs) are malfunctional, causing unavailable user traffic data under their coverage areas. The MADT framework consists of three core modules: 1) DT data construction, 2) DT modeling for traffic prediction, and 3) DT model calibration. For the data construction, we generate a distribution of user volumes over a considered region, where a clustered modular mobility model (clustered-Mo3) is tailored to characterize the post-disaster user movements and a spatial-temporal-aware-k-nearest-neighbors (STAK)-based data imputation technique is applied to supplement the missing user traffic data under malfunctional BSs. For traffic prediction, a spatial-temporal graph convolutional network (STGCN) is utilized to establish the DT model under a sequence-to-sequence (seq2seq) forecasting architecture. To improve the traffic prediction accuracy, we further develop a residual calibration (ResCAL) model to estimate and calibrate the traffic prediction errors. The MADT framework establishes a complete DT lifecycle, where the DT model performance is continuously monitored and fed back to trigger the model update if the network traffic pattern changes. Experimental results show that the MADT framework outperforms state-of-the-art spatiotemporal prediction schemes in terms of prediction accuracy and adaptation to user mobility, achieving performance comparable to that of the complete-data-trained STGCN (CD-STGCN).