Network slicing is a pivotal concept in the evolution of 5G networks. It empowers network operators to divide their physical resources, from the network edge to the data center, concurrently multiplexing multiple tenants while adhering to each of their Service Level Agreements (SLAs). Efficient resource allocation among slices/users with different SLAs is a critical task in this context. The increasing complexity of the problem setup, due to the diversity of services, traffic, SLAs, and network algorithms, makes resource allocation a daunting task for traditional (model based) methods. Consequently, recent research has been prompted towards data-driven methods and Deep Neural Networks (DNNs). Although such methods excel at the application level (e.g., for image classification), applying them to wireless resource allocation is challenging. Not only are the required latencies significantly lower (e.g., for resource block allocation per OFDM frame), but the cost of transferring raw data across the network to centrally process it with a heavy-duty Deep Neural Network (DNN) can also be prohibitive. To address these challenges, Distributed Deep Neural Network (DDNN) architectures have been proposed. These architectures divide DNN layers between the network edge and the central cloud, thereby reducing communication overhead and improving speed. In this setup, a subset of DNN layers at the edge performs initial processing, making quick, localized decisions. If the local processing yields satisfactory results, additional latency and communication costs are avoided; otherwise, intermediate features are transmitted to the central cloud for further processing. There is an intelligent offloading mechanism that delegates a fraction of hard decisions to additional DNN layers situated in the remote cloud when needed. The intelligent offloading mechanism ensures that only the most challenging decisions are delegated to the cloud, optimizing resource usage. To implement offloading, we use (i) a Bayesian confidence-based mechanism that employs dropout during inference to estimate the confidence level of local predictions and (ii) a data-driven module that classifies data samples as either “remote” or “local”. We propose distributed DNN architectures based on CNN and LSTM networks for this task. We investigate (i) the (offline) joint training of DDNNs’ local and remote exits and (ii) optimizing (online) offloading mechanisms to address the challenges of rapid resource scaling and data transmission. Using the publicly available Milano dataset, our experimental results demonstrate that our architecture resolves nearly 50% of decisions at the edge with no additional SLA penalties compared to state-of-the-art centralized models.
Distributed optimization and machine learning for virtualized 6G wireless networks
Thesis
Type:
Thèse
Date:
2024-10-15
Department:
Systèmes de Communication
Eurecom Ref:
7708
Copyright:
© EURECOM. Personal use of this material is permitted. The definitive version of this paper was published in Thesis and is available at :
See also:
PERMALINK : https://www.eurecom.fr/publication/7708