Source-linked AI summary
DRL-driven RAN Slicing Management: A V2X-oriented Approach In Multi-service Scenarios
Daniel E. Garcia-Fernandez, Pablo Vera-Soto, Sergio Fortes, M. Martinez, I. de-la-Bandera, M. L. Luque, A. Mendo, J. Ramiro, Raquel Barco
TL;DR
V2X and other heterogeneous services put demanding requirements on networks, while RAN spectrum scarcity complicates end-to-end SLA compliance. This paper proposes a cell-level PPO framework that dynamically allocates RAN resources between V2X and eMBB slices. On a real 5G SA testbed, it reports V2X delays below the 20 ms threshold alongside 75.7% eMBB efficiency, outperforming static and proportional strategies under the evaluated conditions.
Problem
RAN spectrum scarcity is a bottleneck for end-to-end SLA compliance, and dynamic resource management is complex for heterogeneous services with stringent requirements.
Source: §I · p. 1Method
A cell-level PPO DRL agent dynamically allocates RAN resources, prioritizing V2X SLA fulfillment while also seeking eMBB throughput.
Source: §I · p. 2, §III · p. 3Results
The DRL system reported 90th-percentile reporting-window mean V2X delays of 17.44 and 19.79 ms, both below the 20 ms threshold, and 75.7% 10th-percentile eMBB efficiency; it outperformed static and proportional strategies.
Source: C. Online Performance Validation · p. 7, §V · p. 7Takeaways & Limitations
Under the evaluated conditions, the adaptive policy maintained V2X latency below the configured SLA while achieving high resource efficiency for eMBB tenants.
Source: §V · p. 7
Abstract
from arXiv · showhide
The integration of Vehicle-to-Everything (V2X) communications is driving a profound transformation in vehicular connectivity, expected to significantly enhance traffic efficiency and safety. However, the stringent requirements of V2X services, particularly ultra-low latency and high reliability, present significant technical challenges. 5G's Network Slicing emerges as a key enabler by providing tailored virtual networks that ensure isolation and adaptability for heterogeneous services. This work proposes an intelligent Radio Access Network (RAN) slicing management framework specifically designed for scenarios where safety-critical V2X and high-capacity eMBB slices coexist. In such complex environments, harmonizing conflicting traffic requirements demands continuous, data-driven optimization. To achieve this, the proposed framework leverages an advanced Deep Reinforcement Learning (DRL) approach which dynamically optimizes resource allocation in real time. The framework is empirically validated on a real 5G Standalone (SA) network, where experimental results demonstrate that the DRL-driven approach successfully balances both objectives, outperforming traditional static and proportional allocation strategies by minimizing SLA violations while ensuring high resource utilization for eMBB slices.
I. INTRODUCTION
V2X services bring demanding, heterogeneous requirements to mobile networks, while RAN slicing must adapt scarce radio resources in dynamic conditions. The paper proposes a cell-level PPO framework for coexisting V2X and eMBB slices and evaluates it on a real 5G SA network.
I. INTRODUCTION
- High-mobility, safety-critical V2X services and heterogeneous service requirements challenge networks designed around one-size-fits-all architectures.5G network slicing provides virtual networks tailored to different service needs over shared infrastructure.Source: §I · p. 1
- RAN spectrum scarcity is a bottleneck for end-to-end SLA compliance, and traditional iterative algorithms struggle with dynamic radio conditions.The introduction identifies DRL as a promising way to learn resource-allocation policies through interaction in complex environments.Source: §I · p. 1
- Prior work on coexisting V2X and eMBB slices has lacked explicit strict-SLA enforcement or relied on traditional allocation; real-world deployment evidence remains limited [9].The cited studies are described as predominantly simulation-validated.Source: §I · pp. 1–2
- The proposed cell-level PPO DRL agent prioritizes V2X SLA compliance while maximizing eMBB throughput, and is evaluated in a closed-loop real-world 5G SA deployment.The framework targets safety-critical V2X slices with stringent latency SLAs alongside capacity-hungry eMBB slices.Source: §I · p. 2
II. RL FOR SLICING IN V2X SCENARIOS
This section motivates DRL for dynamic slicing: radio resources change rapidly, making local adaptation important, while PPO offers a stable learning approach for network management.
II. RL FOR SLICING IN V2X SCENARIOS
- Orchestrating diverse resources across network segments and management levels increases slicing complexity, which DRL approaches address through dynamic resource allocation.Source: §II · p. 2
A. Resource Management for Slicing
- Radio resources require more dynamic management than relatively static core configurations, motivating centralized and local control at different timescales.The slice lifecycle includes preparation, commissioning, operation, and decommissioning; during operation, control spans Core, Transport, and RAN segments.Source: A. Resource Management for Slicing · p. 2
B. Reinforcement Learning for Dynamic Allocation
- DRL is presented as an alternative to traditional optimization for the high-dimensional state-action spaces and rapid dynamics of V2X radio environments.PPO is highlighted for stability, ease of implementation, and sample efficiency, which matters when operator datasets are scarce and online training poses practical risks.Source: B. Reinforcement Learning for Dynamic… · pp. 2–3
III. V2X-ORIENTED INTELLIGENT RAN SLICING FRAMEWORK
The framework uses local, cell-level RAN control in three independent layers. Its business layer extracts SLAs and sets tenant priorities, while network partitions are dynamically configured during operation.
III. V2X-ORIENTED INTELLIGENT RAN SLICING FRAMEWORK
- The work proposes a DRL-driven, cell-level intelligent resource-management framework with three independent layers, Business, Network, and Management, to enable robust closed-loop control over the RAN spectrum.Source: §III · p. 3
A. Architecture Layers
- Deployment selects a DRL model based on service count and requirements, followed by a closed operational loop that dynamically configures resource partitions from agent observations.The framework separates Business, Network, and Management layers.Source: A. Architecture Layers · p. 3, §III · p. 3
1) Business Layer:
- The Business Layer extracts tenant SLAs before deployment and establishes strict priorities when congestion makes simultaneous fulfillment of all demands infeasible.Source: 1) Business Layer · p. 3
- For this scenario, V2X latency thresholds take priority; eMBB throughput is maximized after V2X requirements are met, without over-provisioning.Source: 1) Business Layer · p. 3
2) Network Layer:
The network layer controls RAN PRB partitions using soft-slicing quotas and collects telemetry. The management layer processes those observations and selects and runs a suitable RL model in a closed loop.
2) Network Layer:
- The Network Layer focuses on RAN PRB partitioning, using soft-slicing quotas that permit dynamic sharing; minimum quotas set prioritization while maximum quotas are 100% for all slices.Telemetry includes DU physical-resource and channel-quality measurements alongside CU flow statistics.Source: 2) Network Layer · p. 3
B3. State variables
- Cell context for model selection includes current load, interference, active slices, and targeted SLAs; the model is chosen with service count and requirements in view.Source: B3. State variables · p. 4
3) Management Layer:
- The Management Layer’s blocks engineer KPIs, select a pre-trained RL model and scaler, preprocess and scale inputs, then issue PRB quota adjustments in a closed loop.Feature engineering can estimate average RLC PDU delay from cumulative delays and transmitted-PDU counts.Source: B3. State variables · p. 4
B. RL Agent Design
The PPO agent is trained by defining an MDP through its state variables, action space, and reward function, which guides its policy toward the optimization objective.
B. RL Agent Design
- The MDP specifies the environment observation, possible agent actions, and reward function used to guide the PPO policy.The PPO agent operates at the core of the Management layer.Source: B. RL Agent Design · p. 4
- The reward function directs the policy toward the optimization objective.Source: B. RL Agent Design · p. 4
- The PPO agent operates at the core of the Management layer.Source: B. RL Agent Design · p. 4
1) State Variables:
The design uses slice-differentiated performance and cell-level radio metrics as observations, then selects constrained PRB quotas from a discretized set of allocation choices.
1) State Variables:
- State profiles cover V2X latency, eMBB throughput, resource utilization, and channel quality, using counters supported by the commercial network.RLC buffer delay estimates latency because true end-to-end latency is rarely observable from the RAN alone.Source: 1) State Variables · pp. 4–5
- The four-slice scenario uses a 31-feature observation vector: five features per slice and 11 cell-level radio-condition features.Per-slice features include mean PRB utilization, minimum PRB quota, active users, PDCP throughput over the preceding five minutes, and mean RLC delay; cell features summarize CQI and MCS histograms.Source: 1) State Variables · p. 5
2) Action Space:
- At each decision point, the agent selects a minimum PRB quota for every active slice, with the quotas limited by total cell capacity.In the four-slice scenario, quotas are assigned to eMBB-1, eMBB-2, V2X-1, and V2X-2.Source: 2) Action Space · p. 5
- The action space contains 20 predefined quota combinations spanning V2X-prioritized to eMBB-oriented allocations.V2X-1 quotas range from 20% to 2%, V2X-2 from 20% to 4%, eMBB-1 from 23% to 41%, and eMBB-2 from 35% to 62%; eMBB-2 generally receives more because of higher traffic demand.Source: 2) Action Space · p. 5
3) Reward Function:
The reward prioritizes V2X latency-SLA compliance; when all V2X latency SLAs are met, it encourages eMBB throughput while discouraging excessive PRB quotas.
3) Reward Function:
- The reward follows a simulation-based formulation [9], extended through closed-loop validation on commercial 5G equipment, and implements the Business Layer service hierarchy.V2X SLA compliance takes priority; once latency boundaries are satisfied, remaining resources should be dedicated to eMBB services.Source: 3) Reward Function · p. 5
- If any V2X slice exceeds its latency threshold, the reward applies a severe penalty proportional to the violation and ignores eMBB performance.Source: 3) Reward Function · p. 5
- When all V2X latency SLAs are fulfilled, the reward promotes eMBB maximization by penalizing over-provisioning and rewarding eMBB throughput.This branch encourages unused capacity to be repurposed for eMBB rather than granting V2X unrequested performance.Source: 3) Reward Function · p. 5
- A PRB over-provisioning term encourages allocated quotas to track observed utilization, while an eMBB throughput term is scaled relative to per-slice offered rates.The offered rates are 40 and 160 Mbps, and all reward terms are scaled by weights.Source: 3) Reward Function · p. 5
A. Experimental Testbed and Scenario
The evaluation uses a real 5G SA testbed and four slices carrying eMBB and vehicular traffic, with two V2X load configurations and 20 ms latency SLAs.
A. Experimental Testbed and Scenario
- The testbed uses a virtualized Nokia 5G Core and a 5G RAN in n78 at 3.5 GHz, with 50 MHz bandwidth and 30 kHz subcarrier spacing.The DU MAC scheduler supports per-slice dynamic PRB quotas, and PM counters are collected at five-minute intervals.Source: A. Experimental Testbed and Scenario · p. 5
- An Amarisoft UEsimbox emulates up to 64 concurrent users, while an edge server generates traffic and executes the slice-management framework.The framework fetches metrics and configures PRB partitions based on the RL agent's actions.Source: A. Experimental Testbed and Scenario · p. 5
- The four-slice scenario combines persistent eMBB congestion from CBR eMBB-1 and bursty VBR eMBB-2 traffic with V2X traffic patterns drawn from 3GPP TR 37.885.V2X-1 represents dense users with light periodic CAM and DENM traffic; V2X-2 represents fewer users sending larger data volumes.Source: A. Experimental Testbed and Scenario · p. 6
- The two V2X load configurations shift users from 48 to 42 on V2X-1 and from 12 to 18 on V2X-2, with a 20 ms SLA for both slices.V2X-2 experiences increased data density in the second load, while V2X-1 network pressure remains relatively stable.Source: A. Experimental Testbed and Scenario · p. 6
B. Training Phase
Training used empirical telemetry rather than live exploration, then the frozen PPO policy was evaluated on a 5G testbed against two allocation baselines. Across changing loads, the DRL system achieved a joint latency–efficiency compromise while the baselines exposed contrasting trade-offs.
B. Training Phase
- Telemetry covered 20 discrete actions under two load configurations, yielding 1,440 samples; samples within each action-load pair were split 70/30 for training and validation.Each action-load pair was monitored for three hours at five-minute intervals, producing 36 samples.Source: B. Training Phase · p. 6
- The PPO agent trained on alternating Load 1 and Load 2 episodes, stabilized around episode 600, and stopped early around episode 1,600.The training environment was measurement-driven and the load alternated unpredictably to discourage overfitting to one traffic profile.Source: B. Training Phase · p. 6
C. Online Performance Validation
- The trained policy was frozen for a continuous six-hour testbed evaluation as traffic switched between Load 1 and Load 2 every 30 minutes.Equal share and Proportional Share served as static benchmarks, allocating equal quotas or quotas proportional to slice data volume, respectively.Source: C. Online Performance Validation · p. 6
- The eMBB efficiency metric measures how closely observed PRB utilization for both eMBB slices matches their configured minimum quotas; 100% means an exact match.Lower scores indicate larger deviations, whether utilization falls below or exceeds quota.Source: C. Online Performance Validation · p. 6
- Equal share gave V2X excessive permanent resources and kept eMBB efficiency below 10%, whereas Proportional share exceeded 72.2% 10th-percentile efficiency but frequently violated V2X SLAs.V2X-2 under Proportional share showed separate load-level lobes as users increased in Load 2.Source: C. Online Performance Validation · pp. 6–7
- 17.44 and 19.79 ms 90th-percentile reporting-window mean V2X delays, both below 20 ms, accompanied 75.7% 10th-percentile eMBB efficiency under DRL.Equal share drove eMBB efficiency below 10%, while Proportional share had frequent severe SLA violations, especially for V2X-2 under Load 2.Source: C. Online Performance Validation · pp. 6–7, Figure 5 · p. 7
V. CONCLUSIONS AND FUTURE WORK
The paper concludes that its PPO-based RAN slicing framework learned an adaptive policy on a real-world 5G SA testbed, meeting the stated V2X latency target while maintaining eMBB resource efficiency under evaluated conditions. Future work proposes Open RAN deployment and Digital Twin-assisted offline training.
V. CONCLUSIONS AND FUTURE WORK
- Under evaluated conditions, the real-world 5G SA testbed results showed 90th-percentile V2X latencies below the 20 ms SLA alongside high eMBB resource efficiency.The learned adaptive policy outperformed static and proportional allocation strategies.Source: §V · p. 7
- Future work will deploy the agent as an xApp/rApp in the Open RAN RIC to reduce decision latency and increase portability through open interfaces.Source: §V · pp. 7–8
- The authors propose integrating Digital Twins into offline training to bridge simulated exploration and real-world deployment safely.Source: §V · p. 8