RAN Slicing in 5G Networks

• RAN slicing is a technology that partitions a shared RAN into independent virtual slices with tailored performance and SLAs.
• It employs layered resource abstraction and NFV/SDN orchestration, enhanced by AI/ML approaches for dynamic, efficient resource allocation.
• Key applications include 5G enhancements, industrial IoT, vehicular networks, and tactical communications with stringent QoS demands.

Radio Access Network (RAN) slicing is a foundational paradigm in 5G and beyond networks, enabling the partitioning of a common physical RAN infrastructure into multiple, logically isolated, and service-customizable slices. Each slice can be tailored to diverse Service Level Agreements (SLAs), supporting an array of verticals such as eMBB, URLLC, mMTC, and industrial, vehicular, and tactical applications. RAN slicing spans all resource layers—including spectrum, time-frequency resources, compute at edge/MEC servers, and virtualized network functions—under unified orchestration. The following sections systematically examine RAN slicing principles, architectures, performance guarantees, algorithmic frameworks, and emerging research trends, with technical depth and reference to leading arXiv literature.

1. Fundamentals and Deployment Architectures

RAN slicing logically partitions a radio access infrastructure (base stations, spectrum, processing) into self-contained virtual slices, each acting as an independent mobile network for a tenant or service (Saleh et al., 2022). The architecture is layered:

• Physical layer: base stations (gNBs or eNBs), antennas, spectrum, edge/MEC compute, and fronthaul/backhaul.
• Virtual resource layer: abstraction of radio resource blocks (RBs), compute, and storage, typically via NFV and SDN mechanisms.
• Management and Orchestration (MANO): centralized control (e.g., ETSI NFV-MANO, 3GPP NRM) allocates/schedules virtual resources subject to per-slice SLA constraints.

Modern deployments exploit frameworks such as O-RAN with Non-RT RIC (for AI/ML-driven long-term policies) and Near-RT RIC (for fast adaptation), supporting xApps for monitoring and closed-loop slice policy optimization (Cheng et al., 2024, Filali et al., 2022, Filali et al., 10 Jun 2025).

RAN slicing enables:

• Multi-tenancy: multiple MNOs/MVNOs or diverse services on a single RAN.
• SLA isolation and customization: strong per-slice QoS differentiation in throughput, latency, reliability, and resource sharing approaches (static vs. dynamic partitioning).
• Programmability: open interfaces (E2SM-CCC, Protobuf/ASN.1) and software-defined scheduling with adaptive policies (Cheng et al., 2024).

2. Mathematical Models and Performance Guarantees

RAN slicing involves mapping high-level service requirements into resource partitions with explicit performance metrics. Canonical formulations include:

• Operator-to-waveform allocation: Joint optimization across MNO, infrastructure, and time-frequency scheduling levels. Resource assignment variables $x_{i,b,r}$ solve integer or mixed-integer programs maximizing aggregate weighted throughput, demand satisfaction, or RB alignment for CoMP/MIMO (D'Oro et al., 2019).
• Isolation and sharing constraints: Each slice receives dedicated ($R_{i,\min}$) and maximal ($R_{i,\max}$) resource caps, and the resource allocation problem is typically subject to $\sum_{i,j} x_{i,j} \leq R_\mathrm{tot}$ and isolation/priority constraints (Saleh et al., 2022, Zangooei et al., 18 Feb 2026).
• Delay and probabilistic QoS: Stochastic Network Calculus is employed to derive rigorous probabilistic delay bounds for flows/slices, particularly for uRLLC in industrial settings, resulting in delay envelope expressions such as

$$W_f(\theta, \delta) = -2\frac{ \ln(\epsilon_f/2) + \ln(1-e^{-\theta\delta}) }{ \theta (\rho_{S_f}(\theta) - \delta) }$$

subject to stability constraints (Adamuz-Hinojosa et al., 30 Mar 2026).

• Resource multiplexing and burstiness: To fight over-provisioning and underutilization under heavy-tailed (Pareto) bursty traffic, frameworks like HyRA combine minimal per-slice dedicated reservation with a shared pool, balancing statistical multiplexing with delay isolation (Zangooei et al., 18 Feb 2026).
• Physical layer enforcement: The RAN Slicing Enforcement Problem (RSEP) formulates PRB allocation as a binary quadratic program aiming to maximize the number of time-frequency "linked" RBs across interfering BSs for each MVNO, with NP-hardness proven for the general problem. Approximations include convex-relaxations and greedy Most Linked First (MLF) heuristics (D'Oro et al., 2019).

3. Algorithmic Approaches and Scheduling Solutions

A suite of algorithmic frameworks underpin practical RAN slicing:

• Two-stage resource and scheduling frameworks:
  • Outer loop: Optimizes per-slice dedicated + shared resource budgets (e.g. PRBs) via sample average approximation and big-M encoded MIP formulations capturing both SLA constraints and bursty arrivals (Zangooei et al., 18 Feb 2026).
  • Inner loop: Per-TTI or per-slot proportional fair or water-filling schedulers allocate dedicated+shared pools to per-slice and per-UE queues, with stationarity/KKT conditions ensuring optimality under channel state (Zangooei et al., 18 Feb 2026, Cheng et al., 2024).
• AI/ML and DRL-based orchestration:
  • Single-agent DRL: Deep Q-Networks or DDPG optimize per-slice bandwidth shares (Abouaomar et al., 2022, Filali et al., 10 Jun 2025), with state including channel, queue, and SLA metrics, and reward functions encoding SLA adherence, QoS, and resource cost.
  • Hierarchical/two-level DRL: Decouples resource slicing for radio (gNB) and compute (MEC edge), using double DQNs for each level and coordinated policy inference (Filali et al., 2022).
  • Federated and transfer learning: MVNOs train local DRL models for RAN slicing and aggregate globally via FedAvg. Transfer RL allows reuse of expert knowledge (Q-tables or nets) from single-resource tasks to accelerate multi-resource joint slicing (Abouaomar et al., 2022, Zhou et al., 2022).
• Optimization and resource embedding:
  • VNF embedding: RAN slices are virtual network graphs to be mapped on substrate nodes and links. NP-hard formulations are addressed by fast heuristics (resource/degree sorting, group-based packing), as well as DRL-based agents for flexible, high-acceptance embedding under resource constraints (Nguyen et al., 2022, Le et al., 2022).
  • Interference-aware scheduling: Queueing-theoretic models for multi-cell, multi-MVNO systems, with state-dependent service rates and cooperative channel-assignment policies, provide explicit delay/throughput predictability and improved inter-slice isolation (Hashemian et al., 2022).
• Experimental testbeds: SD-RAN controllers (e.g. 5G-EmPOWER, ORANSlice) implement slice-aware admission control (multi-cell, multi-slice), two-level scheduling (inter-slice WRR/PF, intra-slice PF/RR), and 3GPP Release 15–compliant slice lifecycle and monitoring (Koutlia et al., 2019, Cheng et al., 2024).

4. Isolation, SLA Enforcement, and Trade-Offs

RAN slicing must enforce strong isolation and predictability for diverse service SLAs, balancing against efficient resource utilization:

• Granularity of isolation: Options include per-line, per-flow, per-group, and hybrid slicing (e.g. critical flows get dedicated slices, non-critical share pool), with per-flow ensuring the tightest matching of delay targets under resource scarcity (Adamuz-Hinojosa et al., 30 Mar 2026).
• Statistical multiplexing: Pure reservation safeguards strictest SLAs but is resource-expensive; pure sharing improves utilization but risks SLA violations under simultaneous traffic peaks. HyRA and similar hybrids statically assign minimal dedicated budgets and allocate shared pools to absorb rare bursts, achieving 50–75% PRB savings at high compliance (Zangooei et al., 18 Feb 2026).
• Isolation KPIs: Delay and throughput deviations from baseline under interfering load (ADD, VDD, ATD, VTD) precisely quantify inter-slice leakage and robustness (Hashemian et al., 2022).
• Slicing policies and enforcement: Algorithmic solutions (greedy MLF, convex relaxations, AI-driven prioritization) can efficiently enforce owner-defined slicing policies (per-BS, per-MVNO PRB shares) with near-optimal alignment for coordination technologies (e.g., CoMP), with 2× network throughput observed under coordinated enforcement (D'Oro et al., 2019).

5. RAN Slicing for Emerging Applications

Recent research demonstrates the application of RAN slicing to scenarios with stringent requirements:

• Industrial and Time-Critical Applications: In Industry 4.0, per-flow slicing is uniquely capable of guaranteeing sub-millisecond delays for heterogeneous production lines, with SNC-based planners running at non-RT scales on O-RAN platforms (Adamuz-Hinojosa et al., 30 Mar 2026).
• Smart Grid: DRL-based RAN slicing for IEC 61850 services differentiates between slices for real-time protection (hard isolation, ultra-low latency), phasor measurements, and SCADA, embedding service-aware scheduling and functional splits programmable via RANA, RIL, and RSFs (Carrillo et al., 2022).
• Vehicular Networks (IoV): Two-layer constrained DRL (RAWS) combines resource allocation (spectrum, compute) and workload splitting, guaranteeing delay and stability with minimal long-term cost under non-stationary traffic (Wu et al., 2020).
• Tactical Networks: Two-stage DRL (e.g., TD3 for inter- and intra-slice bandwidth allocation) is deployed on O-RAN, emphasizing inter- and intra-slice isolation alongside efficient utilization under adversarial or highly dynamic bandwidth (Filali et al., 10 Jun 2025).
• mIoT and URLLC multiplexing: Joint stochastic queue analysis and SAA+ADMM optimization supports simultaneous bursty access and reliability on shared resources with explicit resource and energy efficiency trade-offs (Yang et al., 2020).

6. Open Research Directions and Implementation Challenges

• Scalability and Complexity: Although hybrid slicing and DRL frameworks achieve strong results in simulations, mixed-integer and neural approaches often hit RT scale limits; further work is needed for scalable, explainable, and certifiable real-time algorithms, particularly for massive device counts and hard SLA slices (Zangooei et al., 18 Feb 2026).
• Dynamic adaptation: Rapidly-varying and non-stationary traffic requires continuous retraining of DRL agents (integrated between non-RT and near-RT RIC), adapting to real-world mobility, traffic surges, and user/service churn (Filali et al., 2022).
• Transfer, federated, and graph learning: Generalizing slice allocation policies across sites, tenants, and services via federated or transfer learning improves robustness and privacy; graph-based models can further encode dynamic network topologies (Zhou et al., 2022, Abouaomar et al., 2022).
• Isolation under interference and multi-cell: Analytical quantification of inter-slice isolation (e.g. via product-form state-dependent queueing) and interference-aware, coordinated channel allocation are active topics, with demonstrated gains in large testbeds (Hashemian et al., 2022).
• Integrated resource orchestration across domains: Coordinating RAN slicing with core/backhaul slicing, fronthaul capacity, and compute/network function placement remains an open multi-domain challenge (Nguyen et al., 2022, Le et al., 2022).

RAN slicing is thus a rapidly evolving, multi-layered field at the intersection of wireless resource allocation, virtual network embedding, AI-driven management, and SLA-driven service orchestration, underpinned by rigorous stochastic and optimization-theoretic models and validated by emerging open-source and experimental testbeds.