Abstract
One of the key benefits of virtualized radio access networks (vRANs) is network management flexibility. However, this versatility raises previously-unseen network management challenges. In this paper, a learning-based zero-touch vRAN orchestration framework (LOFV) is proposed to jointly select the functional splits and allocate the virtualized resources to minimize the long-term management cost. First, testbed measurements of the behaviour between the users' demand and the virtualized resource utilization are collected using a centralized RAN system. The collected data reveals that there are non-linear and non-monotonic relationships between demand and resource utilization. Then, a comprehensive cost model is proposed that takes resource overprovisioning, declined demand, instantiation and reconfiguration into account. Moreover, the proposed cost model also captures different routing and computing costs for each split. Motivated by our measurement insights and cost model, LOFV is developed using a model-free reinforcement learning paradigm. The proposed solution is constructed from a combination of deep Q-learning and a regression-based neural network that maps the network state and users' demand into split and resource control decisions. Our numerical evaluations show that LOFV can offer cost savings by up to 69% of the optimal static policy and 45% of the optimal fully dynamic policy.
Introduction
Modern 5G and beyond networks are undergoing a paradigm shift, moving away from proprietary, hardware-centric architectures toward virtualized Radio Access Networks (vRANs). In a vRAN environment, base station functions are implemented as software components running on generic cloud servers. This transition introduces unprecedented flexibility, allowing operators to implement "functional splits" where the processing tasks of a base station are distributed between a Distributed Unit (vDU) located near the antenna and a Central Unit (vCU) situated in a centralized data center.
While this flexibility is advantageous, it introduces a complex orchestration problem. Operators must simultaneously decide which software tasks to run at which location (the functional split) and how many computing resources (CPU and memory) to allocate to those tasks. Traditional management approaches struggle with this complexity because the relationship between user traffic and resource consumption is often non-linear and non-monotonic. Furthermore, the act of reconfiguring these virtualized resources is not instantaneous or free; it incurs significant overhead and potential service disruptions.
The LOFV Framework
To address these challenges, we propose the Learning-based Zero-touch vRAN Orchestration Framework (LOFV). LOFV utilizes a model-free reinforcement learning (RL) approach to automate orchestration decisions without requiring a predefined mathematical model of the network's internal dynamics. The framework consists of two primary components:
1. Functional Split Orchestration
This component employs a Deep Q-Network (DQN) to determine the optimal functional split. By observing the current network state—including traffic demand, variance, and existing resource levels—the DQN decides whether to maintain the current configuration or trigger a reconfiguration. The agent is trained to balance the trade-off between xHaul bandwidth costs (the connection between vDU and vCU) and the costs associated with centralized vs. distributed computing.
2. Virtualized Resource Orchestration
Once a split is selected, a regression-based neural network maps the traffic demand to specific resource requirements. A critical innovation here is the use of an $\alpha$-OMC loss function. In telecommunications, the cost of under-provisioning (which leads to dropped connections and SLA violations) is significantly higher than the cost of slight over-provisioning (wasted CPU cycles). The $\alpha$-OMC loss function allows the framework to be "cautiously optimistic," prioritizing service availability while still pursuing efficiency.
Empirical Foundation and Methodology
Rather than relying solely on theoretical simulations, LOFV was developed using empirical data from a real-world virtualized Base Station (vBS) prototype. The testbed utilized srsRAN (an open-source 4G/5G suite) containerized within Docker, with Linux Shared Memory emulating the radio interface.
Measurements were collected across varying traffic loads (0 to 35 Mbps) to capture the real-world behavior of CPU utilization. These measurements confirmed that software overhead and inter-process communication create unpredictable resource demands that standard linear models fail to capture. Additionally, the framework explicitly accounts for the "cost of change." In cloud environments, resizing a container or moving a virtual machine can take approximately 23 seconds—a delay that LOFV learns to factor into its long-term cost-benefit analysis.
Performance Evaluation
Numerical evaluations demonstrate that LOFV significantly outperforms traditional orchestration strategies:
- Cost Efficiency: LOFV achieved up to 69% cost savings compared to an optimal static policy (STAO) and 45% savings compared to a fully dynamic policy (DYNO) that fails to account for reconfiguration overhead.
- Convergence and Stability: The RL agent successfully converged to an optimal policy within a few thousand training episodes. It remained robust across various time horizons (2 to 6 hours), indicating its ability to manage both transient traffic spikes and long-term trends.
- Trade-off Management: The framework effectively navigated the "compute vs. bandwidth" trade-off, centralizing functions when bandwidth was cheap and distributing them when local processing was more cost-effective.
Conclusion and Future Directions
LOFV represents a significant step toward zero-touch network management in vRANs. By combining deep reinforcement learning with empirical resource modeling, the framework minimizes management costs while maintaining high service quality.
Future research will focus on scaling this framework to multi-cell environments, where inter-cell interference and shared resource contention introduce additional layers of complexity. Furthermore, we aim to explore energy-aware orchestration to align vRAN operations with global sustainability goals and investigate the portability of these models across different hardware architectures, such as ARM-based servers and specialized hardware accelerators.
Summary updated