Slice resource allocation with distributed deep neural networks for 5G+ networks

End-to-end network slicing is a new concept for 5G+ networks, dividing the network into slices for distinct services. A key task is satisfying service level agreements (SLA) by forecasting how many resources to allocate to each slice. The increasing complexity of the network services makes resource allocation a daunting task for traditional methods. Hence, data-driven methods have been recently explored. Although such methods excel at the application level, their application to wireless resource allocation is challenging. Not only are the latencies required significantly lower, but also the cost of transferring raw data across the network to be processed by a central Deep Neural Network (DNN) can be prohibitive. For this reason Distributed DNN (DDNN) architectures have been considered, where a subset of DNN layers are executed at edge, to improve speed and communication overhead; if it is deemed that a “good enough” allocation have produced locally, the additional latency and communication is avoided; if not, intermediate features produced at the edge are sent to cloud layers. We propose a distributed DNN architecture for this task based on LSTM that excels at forecasting demands with long-term dependencies. We investigate (i) joint training (offline) of the local and remote layers, and (ii) optimizing the (online) decision mechanism for offloading samples either locally or remotely. We show that our architecture resolves nearly 50% of decisions at the edge, with no additional SLA penalty (compared to centralized models).