This repository contains additional data from the paper "Optimal Resource Allocation with Delay Guarantees for Network Slicing in Disaggregated RAN" under review in IEEE/ACM Transactions on Networking. The data from the experiments, the implementation of the optimization model, the simulation parameters and the results obtained are publicly accessible in this repository.

This repository contains the code and experiments related to the article "Optimal Resource Allocation with Delay Guarantees for Network Slicing in Disaggregated RAN".

The work proposes an optimization-based formulation that jointly considers:

• VNF (Virtualized Network Function) placement at RAN nodes (RU, DU, CU).
• Resource allocation in the Transport Network (TN) across network slices.
• End-to-end delay guarantees for URLLC (Ultra-Reliable Low-Latency Communications) use cases in Industry 4.0 scenarios.

The formulation integrates Network Calculus to compute deterministic delay bounds, exploring the Flexible Functional Split (FFS) paradigm while addressing the trade-off between:

• Maximizing the number of flows that meet SLA requirements.
• Minimizing deployment and operational costs of the disaggregated RAN.

To solve the optimization problem, we implement the proposed formulation using the IBM CPLEX CP Optimizer module.

Since Equation (23) contains multilinear and rational terms, the resulting mathematical program is a Non-convex Mixed-Integer Nonlinear Program (MINLP), which cannot be solved directly by the CP Optimizer. Therefore, for the implementation we adopt a conservative (worst-case) latency approximation: every traffic flow is assigned an upper-bound latency estimate and the model is solved with these conservative values. This guarantees that any solution found by the CP Optimizer satisfies the original delay constraints under the assumed worst-case latency scenario (see, 1).

• Different functional splits exhibit clear trade-offs:
  • Split O1 (D-RAN) supports up to 72 UEs but incurs higher costs.
  • Split O9 (C-RAN) has lower costs but supports only 9 UEs under the 1 ms SLA.
• FFS provides flexibility by dynamically adjusting VNF placement to balance cost and performance.

• Comparison among CPLEX (optimal), Reinforcement Learning (DDPG), and Surrogate Model.
• All methods identified Split O6 as the best option, but RL and Surrogate failed to reach the optimal rate allocation.
• CPLEX admitted 118 UEs (100 in one slice + 18 in another), while RL reached only 101 UEs.

• ζ (break-even point): larger values increase DU-related costs, pushing VNF centralization to the CU.
• η (CU/DU cost ratio): higher CU costs reduce the attractiveness of centralization.
• Across all cases, the objective function remained positive, showing feasible and UE-oriented solutions.

• In larger topologies (up to 60 nodes), FFS consistently outperformed fixed options (D-RAN and C-RAN).
• For topologies up to 100 nodes, C-RAN outperformed D-RAN, but FFS remained the most efficient strategy.

Replace the images in /img with the ones generated from your experiments.

• IBM CPLEX: to solve the optimization problem.
• DiscoDNC: to calculate delay bounds via Network Calculus.
• OMNeT++/INET: event-driven network simulations.

• The code and scripts to reproduce the experiments are included in this repository.
• Key results from Section IV of the article can be replicated by configuring the experiments accordingly.

@misc{rocha2023optimal,
      title={Optimal Resource Allocation with Delay Guarantees for Network Slicing in Disaggregated RAN}, 
      author={Flávio G. C. Rocha and Gabriel M. F. de Almeida and Kleber V. Cardoso and Cristiano B. Both and José F. de Rezende},
      year={2023},
      eprint={2305.17321},
      archivePrefix={arXiv},
      primaryClass={cs.NI}
}