Slice Resource Allocation with Multiple Edge-Exit Distributed Deep Neural Networks

Network slicing is a pivotal concept in the evolution of 5 G networks. It enables network operators to partition physical resources from edge to data center, allowing concurrent multiplexing of tenants while adhering to each tenant's Service Level Agreement. Efficient resource allocation is critical in this context, prompting recent research into deep neural networks (DNNs). However, challenges arise with edge resources, including the need to rapidly scale resources (within milliseconds) and the cost of transmitting large volumes of data to a cloud for centralized DNN-based processing. To address these issues, we previously investigated distributed deep neural network (DDNN) architectures based on CNN and LSTM with a single local exit, facilitating efficient edge-cloud collaboration. In this work, we aim to generalize the training methodology for such networks, identifying both shared and unique aspects across different models. Additionally, we propose an extended architecture that incorporates multiple local exits, which introduces new challenges. Unlike DDNNs with one local exit, multiple local exits observe different subsets of the input signals, while the remote exit processes a superset of these. This leads to partially coupled layers and exits, complicating both the model architecture and training process. Moreover, the offloading decision mechanism now involves more intricate trade-offs, such as determining whether to forward specific subsets of preprocessed features to the cloud for further processing. We explore the joint training of DDNN exits and an optimized offloading mechanism, demonstrating that our architecture resolves nearly 40 % of decisions at the edge without incurring additional penalty compared to centralized models.