Optimized Resource and Deep Learning Model Allocation in O-RAN Architecture

In the era of 5G and beyond, telecommunication networks tend to move Radio Access Network (RAN) from centralized architecture to a more distributed architecture for greater interoperability and flexibility. Open RAN (O-RAN) architecture is a paradigm shift that is proposed to enable disaggregation, virtualization, and cloudification of RAN components, possibly offered from multiple vendors, to be connected through open interfaces. Leveraging this O-RAN architecture, Deep Learning (DL) models may be running as a service close to the end users, rather than on the core network, to benefit from reduced latency and bandwidth consumption. If multiple DL models learn on the virtual edge, they will compete for the available communication and computation resources. In this paper, we introduce Optimized Resource and Model Allocation (ORMA), a framework that provides optimized resource allocation for multiple DL models learning at the edge, that aims to maximize the aggregate accuracy while respecting the limited physical resources. Distinguished from related works, ORMA optimizes the learning-related parameters, such as dataset size and number of epochs, as well as the amount of communication and computation resources allocated to each DL model to maximize the aggregate accuracy. Our results show that ORMA consistently outperforms a baseline approach that adopts a fixed, fair resource allocation (FRA) among different DL models, at different total bandwidths and CPU combinations.