Network Slice Load

Slice load is not a single metric but a composite of compute, storage, and transport resource utilization. A slice supporting Ultra-Reliable Low-Latency Communication (URLLC) may be compute-bound on a virtualized Distributed Unit (vDU), while an enhanced Mobile Broadband (eMBB) slice is transport-bound on the N3 interface.

• Compute load: vCPU utilization on virtualized network functions (VNFs) or containerized network functions (CNFs)
• Storage load: Disk I/O and memory consumption for session state and buffering
• Transport load: Fronthaul, midhaul, and backhaul link utilization

Effective predictive balancing must forecast each dimension independently to prevent a single resource bottleneck from violating the slice's SLA.

Each network slice operates under a strict Service Level Agreement (SLA) that defines performance boundaries. Load on one slice must not degrade another, even when they share the same physical gNB. This requires hard isolation or soft isolation mechanisms.

• Deterministic slices: Require guaranteed throughput and bounded latency, often using dedicated PRB allocation
• Non-deterministic slices: Tolerate statistical multiplexing but still have minimum bit rate guarantees

Predictive load balancing must be slice-aware, ensuring that shifting traffic to alleviate congestion in one slice does not violate the resource reservations of a neighboring slice.

Every slice is uniquely identified by its S-NSSAI, which consists of a Slice/Service Type (SST) and an optional Slice Differentiator (SD). This identifier is carried in signaling messages and allows the RAN to apply slice-specific admission control and resource management policies.

• SST values: Standardized types include eMBB (1), URLLC (2), and MIoT (3)
• SD: An optional 24-bit field for distinguishing between multiple slices of the same SST

Predictive models must ingest S-NSSAI as a categorical feature to generate per-slice load forecasts and trigger slice-specific balancing actions.

Slice load is inherently volatile, driven by user mobility, application demands, and time-of-day patterns. When a slice's load approaches its provisioned capacity, the network must either scale the slice vertically (add resources) or trigger inter-slice load redistribution.

• Vertical scaling: Increasing vCPU or PRB allocation within the same physical node
• Horizontal scaling: Instantiating new VNF/CNF instances and redistributing session state

Predictive load forecasting enables proactive scaling, avoiding SLA violations that would occur if scaling were triggered reactively after a threshold breach.

In the RAN, slices are not perfectly isolated at the physical layer. A heavily loaded eMBB slice can cause inter-slice interference that degrades the Signal-to-Interference-plus-Noise Ratio (SINR) of a URLLC slice operating on adjacent PRBs.

• In-band interference: Occurs when slices share the same carrier frequency
• Adjacent channel leakage: Spillover from one slice's transmission into another's frequency allocation

Advanced predictive models incorporate SINR forecasts and interference metrics to anticipate cross-slice degradation and preemptively adjust scheduling weights or power allocation.

Traditional cell-level KPIs like average throughput are insufficient for slice load monitoring. Instead, operators must track per-slice, per-S-NSSAI KPIs to validate SLA compliance.

• Slice throughput: Aggregate downlink and uplink throughput for all PDU sessions within a slice
• Slice latency: End-to-end packet delay measured at the N3 interface per slice
• Slice admission rejections: Count of rejected PDU session establishment requests due to resource exhaustion

These granular metrics serve as both the prediction target for forecasting models and the feedback signal for closed-loop balancing controllers.

The proactive distribution of traffic that respects the isolation boundaries and distinct Service Level Agreements (SLAs) of each network slice. Unlike generic load balancing, this approach forecasts per-slice demand and pre-allocates resources to prevent SLA violations before they occur. Key considerations include:

• Inter-slice resource partitioning: Ensuring one slice's traffic surge doesn't cannibalize another's guaranteed resources
• SLA-driven prioritization: URLLC slices require sub-millisecond latency guarantees, while eMBB slices prioritize throughput
• Slice-specific prediction models: Each slice type exhibits unique traffic patterns requiring tailored forecasting

The specific forecasting of Physical Resource Block (PRB) consumption within a network slice. PRBs are the fundamental unit of time-frequency resource allocation in LTE and 5G NR networks. Accurate prediction enables:

• Proactive admission control for new slice requests
• Dynamic adjustment of slice resource quotas
• Prevention of resource exhaustion in high-demand slices
• Input features typically include historical PRB usage, Channel Quality Indicator (CQI) reports, and active User Equipment (UE) counts per slice

A load distribution strategy that considers the specific Quality of Service (QoS) requirements of different data flows within and across slices. This goes beyond simple load equalization to enforce:

• Guaranteed Bit Rate (GBR) for critical slice services
• Packet Delay Budgets for latency-sensitive slices
• Packet Error Loss Rates for reliability-critical applications
• The balancing algorithm must simultaneously optimize for resource efficiency while maintaining hard QoS commitments defined in each slice's SLA

The implementation of slice-aware predictive load balancing as an xApp running on the Near-Real-Time RAN Intelligent Controller (Near-RT RIC). This architecture enables:

• Control loops operating on 10ms to 1s timescales
• Direct ingestion of per-slice telemetry via the E2 interface
• Closed-loop automation that adjusts slice resource allocation without human intervention
• The xApp hosts the ML inference engine that translates slice load predictions into actionable resource reconfiguration commands

A high-fidelity virtual replica of the sliced RAN environment used to safely train and validate predictive load balancing algorithms. Digital twins enable:

• Testing of extreme slice load scenarios without risking live network stability
• What-if analysis for new slice provisioning and SLA definition
• Accelerated training of reinforcement learning agents for slice orchestration
• Validation of prediction model accuracy across diverse slice configurations before production deployment

A privacy-preserving Federated Learning technique where local slice load prediction models from multiple base stations are averaged on a central server. This approach:

• Creates a robust global model without exposing per-slice user traffic patterns
• Enables collaborative learning across geographically distributed slices
• Preserves the confidentiality of slice-specific resource allocation policies
• The central server aggregates only mathematical model updates, never raw telemetry data