Learning-Based Orchestration for Dynamic Functional Split and Resource Allocation in vRANs

Fahri Wisnu Murti, Samad Ali, George Iosifidis, Matti Latva-aho

TL;DR

Researchers have developed LOFV (Learning-based Zero-touch vRAN Orchestration Framework), an intelligent system that jointly optimizes where cellular network functions are processed and how many computing resources they receive. By using a combination of Deep Q-Learning and regression-based neural networks, the framework can reduce long-term management costs by up to 69% compared to traditional static setups.

What Problem Does This Paper Address?

In modern 5G and beyond networks, we are moving away from proprietary hardware toward virtualized Radio Access Networks (vRANs). In a vRAN, base station functions run as software on generic cloud servers. This offers incredible flexibility: you can split the "brain" of the base station between a Distributed Unit (vDU) near the antenna and a Central Unit (vCU) in a data center.

However, this flexibility creates a massive orchestration headache. Network operators must decide:

  1. The Functional Split: Which specific software tasks run at the vDU vs. the vCU?
  2. Resource Allocation: How much CPU/memory should be assigned to these tasks?

The challenge is that the relationship between user traffic and actual CPU usage is "messy"—it’s non-linear and non-monotonic. Furthermore, changing these settings (reconfiguration) isn't free; it causes delays and consumes power. Doing this manually is impossible at scale, and simple "static" rules lead to massive waste (overprovisioning) or dropped calls (underprovisioning).

Methods

The authors propose LOFV, a two-block framework based on Model-Free Reinforcement Learning (RL):

  • Functional Split Orchestration (The "Where"): This block uses a Deep Q-Network (DQN). It observes the current network state (traffic demand, variance, and previous resource levels) and decides whether to keep the current split or reconfigure to a new one. It learns to balance the cost of xHaul (bandwidth between units) against the cost of computing.
  • Virtualized Resource Orchestration (The "How Much"): Instead of using a simple formula, this block uses a regression-based neural network. It maps the chosen split and traffic demand to specific resource requirements.
  • Asymmetric Cost Modeling: A key innovation is the use of an $\alpha$-OMC loss function. In networking, the "cost" of under-provisioning (dropping a user's connection) is much higher than the cost of slightly over-provisioning (wasting a bit of CPU). This loss function trains the AI to be "cautiously optimistic."

Datasets

The researchers didn't rely on synthetic simulations alone. They built a real-world vBS (virtualized Base Station) prototype using:

  • srsRAN: An open-source 4G/5G software radio suite.
  • Docker: To containerize the network functions.
  • Linux Shared Memory: To emulate the radio interface.

They collected empirical measurements of CPU utilization across different functional splits and traffic loads (0 to 35 Mbps) to create the training data for their resource orchestration model.

Key Insights and Ideas

  • The Non-Linear Reality: Testbed data revealed that CPU usage doesn't always go up just because traffic does. Factors like internal software overhead and inter-process communication make the resource-to-traffic relationship unpredictable using standard mathematical models.
  • The "Cost of Change": The paper explicitly models the "instantiation and reconfiguration cost." Moving a virtual machine or resizing a container takes time (often ~23 seconds in cloud environments), which can disrupt service. LOFV learns when it's better to "stay the course" rather than chase a tiny efficiency gain that is wiped out by reconfiguration overhead.
  • Centralization Trade-offs: Moving functions to the vCU (centralizing) generally reduces total computing costs but spikes the "xHaul" bandwidth requirements. The AI successfully navigates this "compute vs. bandwidth" trade-off.

Key Results

  • Significant Cost Savings: 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 ignores reconfiguration costs.
  • Convergence: The RL agent successfully converged to an optimal policy within a few thousand episodes, even when penalty coefficients for reconfiguration were set very high.
  • Robustness: The framework maintained its performance across different time horizons (from 2 to 6 hours), proving it can handle both short-term spikes and long-term traffic trends.

Limitations and Future Work

  • Single Cell Focus: The current study focuses on a single virtualized base station. In a real-world deployment, an orchestrator would need to manage hundreds of cells simultaneously, introducing complex inter-cell interference and shared resource contention.
  • Hardware Specificity: The resource utilization model was trained on a specific CPU architecture (AMD Ryzen). While the method is general, the specific neural network weights would likely need retraining for different hardware (e.g., ARM-based servers or specialized accelerators).
  • Future Directions: The authors suggest that future work could involve multi-agent RL to handle large-scale deployments and exploring energy-aware orchestration to meet sustainability goals.

Summary updated

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.
BibTeX
@inproceedings{murti2022learning,
  author = {Fahri Wisnu Murti and Samad Ali and George Iosifidis and Matti Latva-aho},
  title = {{Learning-Based Orchestration for Dynamic Functional Split and Resource Allocation in vRANs}},
  booktitle = {EuCNC},
  year = {2022},
  url = {https://openreview.net/forum?id=IHRHCeAS2S},
}

Loading related content...

Figure