Abstract: O-RAN systems and their deployment in virtualized general-purpose computing platforms (O-Cloud) constitute a paradigm shift expected to bring unprecedented performance gains. However, these architectures raise new implementation challenges and threaten to worsen the already-high energy consumption of mobile networks. This paper presents first a series of experiments which assess the O-Cloud's energy costs and their dependency on the servers' hardware, capacity and data traffic properties which, typically, change over time. Next, it proposes a compute policy for assigning the base station data loads to O-Cloud servers in an energy-efficient fashion; and a radio policy that determines at near-real-time the minimum transmission block size for each user so as to avoid unnecessary energy costs. The policies balance energy savings with performance, and ensure that both of them are dispersed fairly across the servers and users, respectively. To cater for the unknown and time-varying parameters affecting the policies, we develop a novel online learning framework with fairness guarantees that apply to the entire operation horizon of the system (long-term fairness). The policies are evaluated using trace-driven simulations and are fully implemented in an O-RAN compatible system where we measure the energy costs and throughput in realistic scenarios.
Introduction
The telecommunications industry is rapidly transitioning toward Open RAN (O-RAN), replacing proprietary hardware with virtualized software running on general-purpose "O-Clouds." While this shift increases flexibility and reduces vendor lock-in, it introduces a critical challenge: energy efficiency. Virtualized base station functions are power-intensive, and their energy costs fluctuate significantly based on radio conditions (SNR) and the specific hardware used (CPUs vs. GPUs). This paper addresses the fundamental tension between energy savings and performance (latency/throughput) while ensuring that resource allocation is fair across all users and infrastructure components.
System Architecture and Control Strategy
The proposed solution utilizes a two-pronged control strategy managed by the RAN Intelligent Controller (RIC) to optimize the O-Cloud environment.
As illustrated in the architecture above, the framework consists of:
- Compute Control (non-Real-Time): This layer decides how to distribute processing workloads across a pool of heterogeneous servers. It dynamically assigns tasks to either CPUs or GPUs based on current demand and efficiency profiles.
- Radio Control (near-Real-Time): This implements a "minimum Transport Block size" (minTB) policy. Because hardware accelerators (HAs) consume a high "base" energy regardless of block size, this policy prevents the transmission of tiny, energy-inefficient blocks, thereby improving energy efficiency per bit.
To manage these policies under volatile conditions, the authors developed an Optimistic Follow The Regularized Leader (OFTRL) framework. This framework uses predictions of future network states to speed up learning and ensures "horizon fairness"—optimizing resources over a long period rather than forcing inefficient millisecond-by-millisecond parity.
Hardware Efficiency and the GPU Paradox
A key contribution of this research is the experimental analysis of how different hardware units handle varying Signal-to-Noise Ratios (SNR).
Experimental data reveals a "GPU Efficiency Paradox": while GPUs process data significantly faster than CPUs (lower latency), they can consume more energy per bit for small workloads or poor radio conditions. As shown in the figure above, the energy cost for GPUs remains relatively high even as SNR improves, whereas CPUs can be more efficient for specific low-load scenarios. This justifies the need for an intelligent compute policy that can switch between processing units based on real-time efficiency metrics.
Key Results and Performance
The framework was evaluated using real-world traffic traces from Frankfurt, Germany, and an O-RAN-compliant testbed featuring Intel Xeon CPUs and NVIDIA V100 GPUs.
- Significant Energy Savings: The system achieved up to 72% energy savings compared to standard "always-on" configurations when users could tolerate an average delay of ~38ms. In more latency-sensitive settings, it still delivered 67% savings with only ~15ms of additional delay.
- Long-Term Fairness: Unlike traditional "utilitarian" algorithms that might starve specific users to maximize total efficiency, the proposed OFTRL framework successfully equalized energy dispersion and performance across the O-Cloud.
- Practical Implementation: The algorithms execute in less than 10ms, meeting the requirements for integration into near-real-time O-RAN controllers.
Conclusion
By combining hardware-aware compute assignment with intelligent radio block sizing, this framework provides a path toward sustainable O-RAN deployments. Future work will focus on tighter integration with AI-based traffic forecasting to further enhance the "optimistic" learning components and scaling the system to support hundreds of users in multi-cell environments.
Summary updated