Fair Resource Allocation in Virtualized O-RAN Platforms

Fatih Aslan, George Iosifidis, Jose A. Ayala-Romero, Andres Garcia-Saavedra, Xavier Costa-Pérez

TL;DR

This paper introduces a novel online learning framework to tackle the high energy consumption of virtualized Open RAN (O-RAN) platforms. By deploying two coordinated policies—one for assigning compute workloads to heterogeneous servers and another for regulating radio transmission sizes—the authors achieve up to 72% energy savings while ensuring long-term fairness for both users and infrastructure providers.

What Problem Does This Paper Address?

The telecommunications industry is moving toward Open RAN (O-RAN), which replaces specialized, proprietary hardware with virtualized software running on general-purpose "O-Clouds" (servers with CPUs and accelerators like GPUs). While this shift increases flexibility, it creates a massive energy problem.

Virtualized base station functions are power-hungry and their energy costs are volatile, fluctuating based on radio conditions (SNR) and the specific hardware used. Furthermore, there is a fundamental tension between saving energy and maintaining performance (latency and throughput). Previous solutions often ignored "fairness"—for example, saving energy by slowing down only a specific group of users or overloading a single server. This paper seeks to balance energy efficiency with performance while distributing the "costs" and "benefits" fairly across the entire system over time.

Methods

The authors propose a two-pronged control strategy managed by the RAN Intelligent Controller (RIC):

  1. Compute Control (non-Real-Time): Decides how to distribute processing workloads across a pool of heterogeneous servers (CPUs vs. GPUs).
  2. Radio Control (near-Real-Time): Implements a "minimum Transport Block size" (minTB) policy. Since hardware accelerators consume nearly the same energy to process a small data block as a large one, the policy prevents the transmission of tiny, energy-inefficient blocks, though this adds some buffering latency.

To manage these policies under unknown and changing conditions, they developed an Optimistic Follow The Regularized Leader (OFTRL) framework. Key technical highlights include:

  • Optimistic Learning: The algorithm uses predictions of future network states to speed up learning. If predictions are perfect, it achieves "constant regret" (optimal performance); if they are wrong, it still maintains robust worst-case guarantees.
  • Horizon Fairness: Instead of trying to be perfectly fair at every millisecond (which is inefficient), the algorithm ensures fairness over a long "horizon," allowing for better short-term resource optimization.
  • Saddle-Point Transformation: They converted complex, non-separable fairness objectives into a manageable mathematical format that allows for efficient, closed-form updates suitable for real-time systems.

Datasets

The study uses a combination of real-world data and a custom-built experimental testbed:

  • Operational RAN Traces: High-resolution traffic and SNR data collected from a live mobile network in Frankfurt, Germany (May 2023).
  • Experimental Testbed: An O-RAN-compliant platform using Intel FlexRAN software, featuring an Intel Xeon Gold CPU and NVIDIA V100 GPUs.
  • Synthetic Scenarios: Used to test the algorithm's robustness against "adversarial" (highly unpredictable) network behavior.

Key Insights and Ideas

  • The GPU Efficiency Paradox: The authors found that while GPUs process data much faster than CPUs, they can consume significantly more energy per bit for small workloads. This justifies the need for a "compute policy" that knows when to use which hardware.
  • Energy-Efficiency of "Bulk" Processing: Hardware accelerators (HAs) have a high "base" energy cost per task. Transmitting larger data blocks (the minTB policy) drastically improves energy efficiency per bit because the HA's power consumption doesn't scale linearly with block size.
  • Fairness as a Multi-Sided Constraint: Fairness isn't just about users. The paper treats fairness as two-sided: ensuring users get similar latency/throughput (performance fairness) and ensuring servers share the energy-reduction burden (cost fairness).

Key Results

  • Massive Energy Savings: In testbed experiments, the system achieved 72% energy savings compared to standard "always-on" configurations, provided users could tolerate an average additional delay of ~38ms.
  • Balanced Trade-offs: For more latency-sensitive scenarios, they demonstrated 67% energy savings with only ~15ms of additional delay.
  • Superior Fairness: Compared to "utilitarian" algorithms (which only maximize total efficiency), the proposed framework successfully prevented "starving" specific servers or users, equalizing energy dispersion across the O-Cloud.
  • Low Overhead: The algorithms execute in less than 10ms, making them practical for integration into near-real-time O-RAN controllers.

Limitations and Future Work

  • Function Normalization: The authors note that operators must carefully normalize utility (performance) and cost (energy) functions to reach the desired balance, which may require manual tuning in different deployment environments.
  • Prediction Dependency: While the "optimistic" aspect helps, the performance gains are significantly higher with accurate traffic predictions. Future work could involve tighter integration with advanced AI-based traffic forecasting.
  • Scaling: While tested with 10 users per vBS in simulations, real-world scaling to hundreds of users per cell in a fully loaded multi-cell environment remains a target for future large-scale field trials.

Summary updated

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.
BibTeX
@inproceedings{aslan2024fair,
  author = {Fatih Aslan and George Iosifidis and Jose A. Ayala-Romero and Andres Garcia-Saavedra and Xavier Costa-Pérez},
  title = {{Fair Resource Allocation in Virtualized O-RAN Platforms}},
  booktitle = {SIGMETRICS/Performance (Abstracts)},
  year = {2024},
  doi = {10.1145/3652963.3655054},
}

Loading related content...

Figure