Odin: Effective End-to-End SLA Decomposition for 5G/6G Network Slicing via Online Learning

Duo Cheng, Ramanujan K. Sheshadri, Ahan Kak, Nakjung Choi, Xingyu Zhou, Bo Ji

Abstract

In the 5G/6G era, network slicing enables the creation of virtual networks tailored to specific service requirements. However, ensuring End-to-End (E2E) Service Level Agreements (SLAs) across heterogeneous and dynamic domains—Radio Access Network (RAN), Transport, and Core—remains a significant challenge. This paper introduces Odin, a framework that leverages Constrained Bayesian Optimization to solve the SLA decomposition problem. By treating network domains as black boxes and utilizing online feedback, Odin achieves up to 45% higher SLA satisfaction and improved resource efficiency compared to traditional heuristic or offline methods.

Introduction: The SLA Decomposition Challenge

Network slicing allows operators to support diverse applications, from ultra-reliable low-latency communications (URLLC) for autonomous vehicles to high-bandwidth streaming. To meet an E2E SLA, such as a 100ms latency target, an E2E orchestrator must partition this "budget" among the RAN, Transport, and Core domains. This process, known as SLA decomposition, is complicated by three primary factors: heterogeneity of hardware, dynamic network conditions, and the opacity of domain-specific resource management.

The SLA decomposition problem and impact on performance.

Traditional solutions often rely on static, offline models that fail to adapt to these dynamic environments, leading to either SLA violations or excessive resource over-provisioning.

The Odin Framework

Odin addresses these challenges through an online learning approach based on Constrained Bayesian Optimization (BO). It treats each network domain as a black box, requiring no knowledge of internal mechanics, which makes it ideal for multi-vendor environments.

Technical Architecture

The core of Odin resides within the Network Slice Management Function (NSMF). It uses Gaussian Processes (GPs) as surrogate models to predict domain performance (e.g., latency) and the associated resource costs for any given decomposition.

The Odin technical architecture.

Key features of the framework include:

  • Online Feedback Loop: Odin proposes a decomposition, observes real-world performance from the domains, and updates its GP models in real-time.
  • Exploration vs. Exploitation: The algorithm balances trying new configurations to learn (exploration) with selecting known efficient ones to minimize costs (exploitation).
  • Sampling Heterogeneity: Odin can update its models for volatile domains (like the RAN) more frequently than for stable ones (like the Core), optimizing communication overhead.
  • Theoretical Guarantees: Odin is backed by a proof of sub-linear regret, ensuring that its decision-making converges toward the optimal strategy over time.

Evaluation and Key Results

The researchers evaluated Odin using a combination of real-world traces from an Open RAN testbed and large-scale simulations.

  • SLA Satisfaction: Odin demonstrated up to 45% higher SLA satisfaction than "Explore-then-commit" (ETC) baselines, particularly in noisy environments.
  • Resource Efficiency: By finding the most cost-effective way to meet targets, Odin significantly reduced the consumption of Physical Resource Blocks (PRBs) in the RAN and CPU/bandwidth in other domains.
  • Robustness: The system remained effective even with feedback noise levels up to 50%, successfully shifting latency budgets as radio conditions fluctuated.

Conclusion

Odin represents a significant step forward in multi-domain network orchestration. By combining the flexibility of black-box modeling with the rigor of Bayesian Optimization, it provides a practical and theoretically sound solution for 5G/6G SLA management. Future work will explore extending Odin to handle multi-objective SLAs, such as simultaneous optimization of latency, jitter, and packet loss.

Summary updated

BibTeX
@inproceedings{cheng2025odin,
  author = {Duo Cheng and Ramanujan K. Sheshadri and Ahan Kak and Nakjung Choi and Xingyu Zhou and Bo Ji},
  title = {{Odin: Effective End-to-End SLA Decomposition for 5G/6G Network Slicing via Online Learning}},
  booktitle = {MobiHoc},
  year = {2025},
  doi = {10.1145/3704413.3764437},
}
Figure