PaperlayerView exact revision on arXiv

Source-linked AI summary

TriSLA: A Preventive and Closed-Loop SLA-Aware Architecture for Multidomain Decision-Making with Explainable Artificial Intelligence in 5G Networks

Abel J. R. Lisboa, Gustavo Z. Bruno, Cristiano B. Both

arXiv:2609.01293v1cs.NI

TL;DR

Multidomain 5G slicing makes SLA assurance difficult because heterogeneous RAN, TN, and 5GC conditions vary dynamically. TriSLA combines ontology-driven intent interpretation, predictive feasibility inference, XAI attribution, and closed-loop runtime assurance, achieving 100% SLA satisfaction for admitted slices in its prototype evaluation.

  • Problem

    Dynamic resource variability and heterogeneous service requirements make end-to-end SLA assurance difficult in multidomain 5G slicing.

  • Method

    TriSLA integrates ontology-driven semantic interpretation, multidomain predictive feasibility assessment, XAI decision attribution, and closed-loop runtime supervision.

  • Results

    100% SLA satisfaction was achieved for admitted slices, while feasibility classification reached 98.68% accuracy with the default Random Forest classifier.

  • Takeaways & Limitations

    TriSLA provides integrated preventive admission and continuous runtime assurance across multidomain 5G slicing.

Abstract

from arXiv · show

Network slicing in multidomain 5G environments introduces critical challenges in guaranteeing Service Level Agreements (SLAs) under dynamic resource variability and heterogeneous service requirements. This article presents TriSLA, a closed-loop, preventive, SLA-aware architecture designed to evaluate feasibility at request time and continuously ensure SLA compliance during operation. The architecture combines ontology-driven semantic intent interpretation, multidomain machine learning feasibility risk inference, Explainable Artificial Intelligence (XAI) feature attribution, and closed-loop runtime SLA assurance into a unified operational pipeline. A fully operational prototype was evaluated in a multi-node cloud-native environment integrating Radio Access Network (RAN), Transport Network (TN), and 5G Core (5GC) domains with real-time telemetry collection. Experimental evaluation demonstrates that TriSLA guarantees a 100% SLA satisfaction rate for admitted slices, completely eliminating post-deployment violations compared to reactive (51.2%) and static threshold (80.4%) admission baselines. The predictive feasibility assessment achieved a classification accuracy of up to 99.51% (98.68% for the default explainable Random Forest classifier), enabling preventive admission decisions before infrastructure commitment. Furthermore, the cognitive admission pipeline introduces minimal processing overhead, requiring 25.37 ms for ontology-driven semantic parsing and 231.66 ms for XAI-assisted feasibility inference. Concurrently, the closed-loop assurance engine resolves 100% of runtime telemetry anomalies within a 4.22 s recovery cycle. These results demonstrate that TriSLA provides reliable, explainable, transparent, and preventive SLA management through integrated predictive admission and closed-loop runtime assurance for next-generation 5G networks.

I. INTRODUCTION

TriSLA addresses multidomain 5G SLA assurance challenges by combining preventive feasibility assessment, explainable decision-making, and closed-loop runtime supervision. Its prototype validates this integrated workflow across RAN, TN, and 5GC domains.

  • Problem: Dynamic interactions across RAN, TN, and 5GC domains make end-to-end SLA consistency difficult to guarantee.Changing latency, allocation, congestion, and stability conditions can compromise guarantees even when individual domains appear adequately provisioned.
  • Problem: Reactive SLA workflows detect violations after degradation and cannot reliably prevent infeasible slice deployments at admission time.Delayed reactions are especially problematic for mission-critical services requiring sustainable SLAs.
  • TriSLA approach: TriSLA unifies ontology-driven intent interpretation, predictive multidomain feasibility inference, XAI attribution, and runtime closed-loop remediation.The architecture evaluates feasibility before instantiation while maintaining alignment with runtime conditions.
  • Evaluation: 98.68% accuracy and 100% SLA satisfaction for admitted slices were achieved by the default Random Forest classifier and preventive filtering.Reactive and static-capacity baselines reached 51.2% and 80.4% SLA satisfaction, respectively.
  • Evaluation: 25.37 ms semantic processing, 231.66 ms ML/XAI inference, and a 4.22 s remediation cycle indicate low admission overhead and rapid anomaly recovery.The prototype operated in a cloud-native simulated environment spanning RAN, TN, and 5GC emulators.

II. THEORETICAL BACKGROUND

The theoretical background frames SLA assurance as coordinated control of heterogeneous RAN, TN, and 5GC resources under dynamic conditions. It motivates preventive management because reactive enforcement occurs only after degradation.

  • 5G slicing: Network slicing partitions shared physical infrastructure into logical networks supporting URLLC, eMBB, and mMTC requirements.Distinct profiles isolate resources and customize traffic-management policies for tenants.
  • Multidomain assurance: Disaggregated and virtualized RAN architectures improve deployment flexibility but increase the complexity of end-to-end service management.Consistent compliance requires coordinated control across heterogeneous RAN, TN, and 5GC infrastructures.
  • SLA foundations: SLAs specify performance, availability, and quality metrics including latency, jitter, throughput, packet loss, and resource utilization.Keeping these indicators within negotiated thresholds is critical for service quality and provider credibility.
  • Preventive management: Traditional frameworks detect violations and initiate corrective actions only after service degradation has occurred.This approach is insufficient when resource availability must be guaranteed at slice instantiation under dynamic multidomain conditions.

C. Explainable Artificial Intelligence for Network Decision-Making

Explainable AI provides interpretable feature contributions for multidomain admission decisions, while preventive closed-loop management connects semantic intake, feasibility evaluation, provisioning, and runtime supervision. This combination addresses opacity and delayed reaction in SLA control.

  • AI decision-making: Telemetry-driven ML can estimate service feasibility before physical infrastructure resources are committed.The models correlate resource metrics with requested SLA thresholds.
  • Explainability: Black-box AI decisions reduce operational trust and hamper troubleshooting during service degradation.Operators therefore require interpretable methods before applying automated decisions to infrastructure.
  • Explainability: XAI translates raw network metrics into explicit contributors to feasibility decisions across RAN, TN, and 5GC conditions.Feature-importance analysis and contribution tracking expose variables affecting predictions.
  • Closed-loop model: The preventive model evaluates feasibility before instantiation and routes requests to provisioning, renegotiation, or rejection.Runtime telemetry feeds back into predictive admission, forming a closed loop.
  • Research gap: Maintaining causal relationships among preventive admission, orchestration, telemetry, and runtime validation remains a significant challenge.Existing platforms often fail to integrate these functions in one framework.

III. RELATED WORK

Prior work covers standards, semantic modeling, AI admission, SLA orchestration, and multidomain platforms, but these capabilities remain fragmented. TriSLA is positioned as an integrated preventive architecture bridging admission control and runtime assurance.

  • Literature landscape: The literature review organizes prior work into standards and surveys, slice admission, SLA management, and multidomain orchestration.This categorization is used to identify operational gaps addressed by TriSLA.
  • Comparison criteria: Table I compares works across preventive admission, multidomain coverage, AI, XAI, semantic modeling, closed-loop execution, and testbed validation.Partial validation support denotes simulated or emulated prototypes rather than commercial hardware.
  • Research gaps: Existing works cover only subsets of the combined requirements, often relying on opaque logic or isolated simulations.These limitations can omit complex multidomain interactions and container-orchestration overheads.
  • TriSLA position: TriSLA coordinates ontology-assisted interpretation, explainable ML inference, real-time telemetry correlation, and closed-loop execution before resource allocation.Its cloud-native prototype connects pre-deployment admission with continuous runtime assurance.

IV. TRISLA ARCHITECTURE

TriSLA is a closed-loop preventive architecture that evaluates slice feasibility before infrastructure commitment and connects semantic requests to multidomain operational configurations. Its three-tier design combines predictive admission, coordinated execution, and runtime assurance across the RAN, TN, and 5GC.

  • TriSLA evaluates slice feasibility before infrastructure commitment and connects semantic intent requests to operational configurations across the RAN, TN, and 5GC.
  • The architecture comprises Intelligence, Execution, and Observability and Runtime Assurance layers interconnected through standardized interface contracts.
  • Intelligence Layer: The Intelligence Layer performs semantic normalization, predictive admission inference, and multidomain feasibility arbitration through the SLA Intake Gateway, SEM-CSMF, ML-NSMF, and Decision Engine.
  • Execution Layer: The Execution Layer translates validated admission decisions into coordinated provisioning actions across RAN, TN, and 5GC domain controllers.
  • Admission decision: The Decision Engine assesses combined multidomain state and returns ACCEPT, RENEGOTIATE, or REJECT outcomes before resource commitment.
  • Explainability: SHAP feature attributions execute asynchronously, preserving explainability metadata without delaying slice instantiation.

C. Observability and Runtime Assurance Layer

TriSLA combines preventive admission and orchestration with continuous runtime supervision. Telemetry is correlated across multidomain slices so the SLA-Agent can detect drift and trigger closed-loop remediation.

  • TriSLA coordinates slice governance through preventive admission and orchestration followed by continuous runtime assurance across five sequential phases.
  • Preventive Admission and Orchestration: Semantic intake normalizes tenant intent into a canonical NEST profile and retrieves an infrastructure telemetry snapshot for evaluation.
  • Preventive Admission and Orchestration: Predictive decision and feasibility arbitration estimates violation risks and confidence, then produces ACCEPT, RENEGOTIATE, or REJECT outcomes while XAI attributions run in background tasks.
  • Preventive Admission and Orchestration: Multidomain provisioning decomposes admitted slice intents into atomic configurations across RAN, TN, and 5GC controllers, then registers runtime supervision parameters.
  • Observability Context Binding: Observability context binding correlates active slice identifiers and telemetry tracking across the RAN, TN, and 5GC domains.
  • Continuous Runtime Assurance: The SLA-Agent executes an Observe-Analyze-Decide cycle to detect drift, contention, and compliance anomalies, publishing alerts that drive automated remediation.
  • By coupling preventive ML admission with continuous observability, TriSLA addresses SLA degradation before commitment while adapting to operational drift.

V. PROTOTYPE IMPLEMENTATION

The TriSLA prototype implements the architecture as containerized Kubernetes microservices with separated admission, orchestration, and telemetry functions. Standardized interfaces support multidomain provisioning and runtime supervision across emulated 5G components.

  • The prototype deploys the SLA Intake Gateway, SEM-CSMF, ML-NSMF, Decision Engine, NASP Adapter, and SLA-Agent as containerized Kubernetes microservices.
  • Dedicated Kubernetes pods, deployments, services, namespace policies, ConfigMaps, and resource quotas provide isolated lifecycle management and configurable runtimes.
  • Intelligence Layer: The SLA Intake Gateway validates JSON requests, while the SEM-CSMF converts quantitative thresholds and semantic descriptors into canonical NEST profiles.
  • Control-plane services communicate through synchronous REST APIs carrying structured JSON over HTTP, with Kubernetes ClusterIP handling service discovery and routing.
  • Helm charts package releases and inject parameters and compute quotas, allowing individual microservices to be updated or scaled independently.
  • The implementation separates admission evaluation, resource orchestration, and telemetry supervision through standardized REST interfaces to preserve control-plane responsiveness.
  • Intelligence Layer: The ML-NSMF uses a scikit-learn Random Forest classifier to evaluate admission feasibility from multidomain telemetry snapshots.
  • Execution Layer: The NASP Adapter decomposes end-to-end intents into domain-specific directives and dispatches provisioning requests asynchronously to RAN, TN, and 5GC controllers.

VI. EVALUATION METHODOLOGY

The evaluation uses an unchanged cloud-native Kubernetes deployment to assess TriSLA across the complete service lifecycle. A consolidated dataset combines semantic, telemetry, admission, runtime, and workflow records collected under controlled multidomain conditions.

  • The study evaluates ontology-driven interpretation, ML feasibility inference, XAI attribution, preventive admission, runtime assurance, and end-to-end processing on an operational prototype.
  • All experiments retain a fixed software, telemetry, and policy configuration throughout evaluation.
  • A single consolidated dataset integrates semantic processing, multidomain telemetry, admission decisions, runtime assurance events, and workflow measurements.
  • Multidomain Emulation Testbed: The testbed uses a cloud-simulated NASP platform across five isolated Kubernetes namespaces on a multi-node cluster.
  • 5G Core: The 5GC namespace deploys free5GC network functions and exposes standard SBI endpoints for automated slice provisioning.
  • RAN: The UERANSIM namespace emulates gNB and UE workloads that generate traffic matching service profiles such as eMBB and URLLC.
  • Transport Network: The transport namespace combines Mininet and ONOS to configure programmable OpenFlow switches, isolated paths, bandwidth, priorities, and link delays.
  • Observability: The monitoring namespace deploys Prometheus, Grafana, OpenTelemetry, and Jaeger to aggregate multidomain metrics and distributed traces for SLA supervision.

B. Multidomain Telemetry and Workload Generation

The evaluation uses multidomain telemetry and controlled workload scenarios to test TriSLA across semantic processing, feasibility inference, admission control, runtime assurance, and lifecycle performance.

  • Telemetry Collection: Telemetry snapshots from the RAN, TN, and 5GC domains were collected before each admission request and linked to decisions and runtime records.Metrics included PRB utilization, TN latency, packet loss, jitter, and 5GC CPU and memory utilization.
  • Workload Generation: Eight controlled scenarios, C0–C7, represented nominal and stressed multidomain operating conditions.Domain-specific generators created RAN connection bursts, 5GC CPU and memory stress, and TN variations.
  • Scenario Design: Compound scenarios C4–C7 introduced concurrent degradation across two or three domains to test semantic resilience and predictive feasibility boundaries.Runtime assurance instead used nominal and single-domain deviation scenarios C0–C3 to isolate recovery dynamics.
  • Evaluation Scope: The evaluation assessed semantic latency and robustness, ML feasibility inference, preventive admission, runtime assurance, and end-to-end lifecycle performance.Admission and runtime measurements included decision outcomes, SLA satisfaction, anomaly detection, recovery, revalidation, and cumulative latency.

A. Semantic SLA Processing Latency and Robustness

TriSLA’s semantic engine converts natural-language SLA requirements into standardized slice templates with low latency while maintaining classification and attribute consistency across operating scenarios.

  • Processing Latency: 25.37±3.38 ms is the average end-to-end latency for ontology-validated semantic transformation into machine-enforceable slice templates.The transformation executes synchronously during online admission.
  • Classification Robustness: 1.00 ± 0.00 is the maximum slice-type classification accuracy and macro F1-score across C0–C7.Requests were correctly categorized as URLLC, eMBB, or mMTC.
  • Semantic Robustness: Attribute consistency ranges from 92.0% in C0 and C3 to 72.0% in C6 and C7.The engine continues extracting individual QoS constraints from multi-clause natural-language specifications.
  • Scenario-Specific Results: 20.0% canonical mapping and 84.0% attribute consistency were achieved for URLLC requests in C7.URLLC benefited from concise latency bounds under compound stress.
  • Fallback Mapping: 0.0% direct canonical mapping occurred for eMBB and mMTC in C7, with attribute consistency of 68.0% and 64.0%, respectively.Deterministic ontology fallback injects defaults for unmentioned non-critical fields while preserving extracted QoS constraints.
  • Feasibility Inference: 99.51% ± 0.33% accuracy was achieved by XGBoost, while the default Random Forest reached 98.68%±0.48%.ML models were trained and validated on telemetry from all eight operating scenarios.

C. Computational Latency of Feasibility Inference and XAI Attribution

TriSLA combines predictive feasibility inference, XAI attribution, preventive admission filtering, and runtime assurance to evaluate and protect slices under multidomain conditions.

  • Inference and XAI: 231.66 ms is the cumulative processing time for predictive admission and SHAP-based XAI feature attribution.Telemetry scaling and normalization take 4.07 ms, followed by predictive model execution at 109.00 ms.
  • Preventive Admission: 112 infeasible requests, or 46.67%, were preventively rejected, while five borderline requests, or 2.08%, were renegotiated.TriSLA evaluates multidomain feasibility before resource allocation.
  • Preventive Admission: 100.0% SLA satisfaction was achieved for all 123 directly admitted slices under TriSLA.The admission evaluation used 240 requests and linked admission filtering to runtime compliance.
  • Baseline Comparison: 80.4% SLA satisfaction was obtained by the static threshold baseline, compared with 51.2% for the reactive always-accept baseline.Static thresholds admitted 153 slices, while the reactive strategy admitted all 240 requests.
  • Runtime Assurance: Closed-Loop and Monitor-Only modes had statistically equivalent detection latency: 2090 ± 3 ms and 2086 ± 4 ms, respectively.The comparison separates telemetry ingestion and threshold evaluation from active remediation.
  • Runtime Assurance: 4216 ms, 4223 ms, and 4241 ms were the closed-loop cycle durations under RAN, TN, and 5GC stress, respectively.Under nominal conditions, the SLA-Agent observed zero false-positive anomaly detections.

F. Detailed End-to-End Admission Latency Analysis

The integrated admission workflow reaches a mean 4046.3 ± 736.5 ms end-to-end latency, with multidomain provisioning dominating the latency budget.

  • End-to-End Latency: 4046.3 ± 736.5 ms is the mean End-to-End admission latency across five sequential macro steps, M01–M05.The workflow integrates semantic intake, predictive arbitration, multidomain provisioning, and observability binding.
  • Macro-Step Breakdown: 110.7 ± 10.4 ms is the SLA Intake and Semantic Processing latency for M01.M01 combines intent ingestion, validation, and profiling in the SEM-CSMF.
  • Macro-Step Breakdown: 936.3 ± 64.6 ms is the Predictive Feasibility and Decision latency for M02.The ML-NSMF evaluates multidomain resource availability and triggers XAI attribution.
  • Latency Dominance: 2.74 s ±720.3 ms is the Multidomain Resource Provisioning latency for M03, the dominant component of the total budget.M01, M02, and observability context binding execute in sub-second durations by comparison.
  • Architectural Integration: TriSLA unifies ontology-driven intent translation, predictive feasibility inference, XAI attribution, and continuous runtime supervision across the service lifecycle.The architecture is presented as a preventive and closed-loop SLA-aware workflow for multidomain 5G environments.
  • Scope and Future Work: Future evaluation will extend beyond the operational testbed to larger deployments, higher request rates, and more heterogeneous network topologies.Planned research includes adaptive policy refinement, online retraining under concept drift, and reinforcement learning for closed-loop control.
Loading 2609.01293v1…