arXiv CSopen access

SLA-Aware Traffic Steering in Hybrid TN-NTN 5G Backhaul: A Potential Game Approach

SLA-Aware Traffic Steering in Hybrid TN-NTN 5G Backhaul: A Potential Game Approach Hojjat Navidan∗ , Delia Rico† , Mohammad Cheraghinia∗ , Ingrid Moerman∗ , Adnan Shahid∗ ∗ IDLab, Department of Information Technology, Ghent University – imec, Belgium

arXiv:2605.26673v1 [cs.NI] 26 May 2026

† ITIS Software, University of Malaga, Spain

Abstract—The integration of Non-Terrestrial Networks (NTN) with Terrestrial Networks (TN) is a key enabler for resilient 5G-Advanced and future 6G backhaul infrastructures. However, managing traffic across these highly asymmetric links remains a significant routing challenge, as systems must support heterogeneous network slices with conflicting service-level agreements (SLAs) while selectively utilizing costly NTN resources. This paper presents a computationally lightweight SLA-aware traffic-steering framework for a hybrid TN-NTN backhaul that models the load-balancing problem as an exact potential game. This mathematical foundation inherently enables decentralized coordination between uplink and downlink load-balancing agents without control-message overhead. By formulating traffic steering as a coupled optimization problem, per-slice (or per-user group) traffic fractions are dynamically distributed across terrestrial and satellite paths based on utility functions that capture throughput, latency, packet loss, and SLA penalties. The resulting game admits a pure Nash equilibrium, ensuring stable and predictable traffic adaptation under non-stationary load conditions. The framework is evaluated on a geographically distributed 5G testbed, using bidirectional traffic generated for five representative slices. Experimental results show that the proposed controller significantly outperforms heuristic and conventional baselines, reducing SLA violations to 1.7% for V2X and 0.7% for the emergency slice while completely eliminating them for video, IoT, and best-effort traffic. Index Terms—Load balancing, 5G, 6G, Non-Terrestrial Networks, Terrestrial Networks, Backhaul, Game Theory

I. I NTRODUCTION The evolution toward 5G-Advanced and emerging 6G network architectures relies heavily on the seamless integration of Non-Terrestrial Networks (NTN), such as Low Earth Orbit (LEO) satellite constellations, with traditional Terrestrial Networks (TN) [1], [2]. This TN-NTN hybrid infrastructure is essential to ensure global coverage and network resilience during outages [3]. Augmenting fiber-optic backhauls with satellite links enables network operators to dynamically pool capacity to meet the increasing bandwidth demands of nextgeneration mobile applications [4]. Furthermore, the 5G network-slicing paradigm requires this unified backhaul infrastructure to support highly heterogeThis work was supported by the Horizon Europe MCSA Staff Exchanges 2021 program under grant agreement 101086218 (EVOLVE); the Smart Networks and Services Joint Undertaking (SNS JU) under the EU Horizon Europe research and innovation program under Grant Agreements No. 101139194 (6G-XCEL project) and No. 101096328 (6G-SANDBOX project); and the Spanish national project LearnFDT under grant agreement PID2022142181OB-I00.

neous traffic classes with strictly enforced, yet conflicting, Service-Level Agreements (SLAs) [5]. These SLA constraints span a diverse range of network slices, from latency-sensitive Vehicle-to-Everything (V2X) and emergency services, to massive Internet of Things (IoT) slices designed for delay-tolerant data bursts. Satisfying these diverse Quality of Service (QoS) requirements demands that the network edge possess the intelligence to seamlessly steer per-User Equipment (UE) traffic flows across available backhaul paths. This will ultimately maximize the overall network performance in terms of throughput, latency, and packet loss, ensuring a consistent user experience without starving any individual slice [6]. However, managing traffic across a hybrid TN-NTN architecture poses complex routing challenges due to severe link asymmetry and economic limitations [6]. While terrestrial fiber links offer low latency and high capacity, NTN links inherently suffer from substantial propagation delays and highly variable channel conditions [7]. Importantly, using NTN resources typically results in significantly higher operational costs due to the inherent expense of satellite bandwidth and the complex infrastructure required for maintenance [8]. Therefore, satellite resources should not be used indiscriminately for all traffic, but rather employed selectively when terrestrial backhaul becomes congested or when resilience is required [9]. This calls for an intelligent load-balancing and traffic-steering mechanism that dynamically accounts for SLA requirements while balancing the latency penalties and resource costs associated with satellite transport. Traditional static load-balancing strategies fail to adapt to the highly asymmetric and cost-sensitive dynamics of hybrid TN-NTN links, risking severe QoS degradation and SLA violations under fluctuating loads. We propose a computationally lightweight, game-theoretic, proactive traffic steering mechanism that dynamically steers per-slice (or per-UE group) traffic across asymmetric TN and NTN backhaul paths. By incentivizing latency-tolerant traffic to offload to the NTN link during heavy load events, the proposed system proactively reserves the terrestrial path for mission-critical slices. The core contributions of this paper are fourfold: •

We formulate the per-slice traffic steering problem as an exact potential game, mapping competing SLAs into utility functions governed by a shared link-capacity penalty.

We design a proactive, decentralized routing controller to solve the game. This computationally lightweight approach anticipates congestion and mathematically guarantees rapid convergence to a pure Nash equilibrium. • We empirically validate the proposed framework on a live, geographically distributed 5G testbed, utilizing commercial Starlink NTN and terrestrial fiber links. • We demonstrate that our approach outperforms static and reactive baselines, reducing SLA violations for priority slices under non-stationary traffic. •

II. R ELATED W ORK Integration of satellites into terrestrial backhauls has become a major research direction for beyond-5G and 6G systems [1]. Recent surveys highlight that hybrid TN-NTN architectures introduce substantial challenges in routing, resource allocation, and QoS assurance due to heterogeneous propagation delays, capacity, and channel dynamics [10]. In multi-RAT and 5G environments, distributed and centralized load-balancing schemes have been proposed to address QoS requirements and uneven traffic distribution [11]. Similarly, 5G access traffic steering, switching, and splitting mechanisms split flows across heterogeneous access points based on utility and service requirements [12]. While these approaches demonstrate the value of adaptive traffic steering, they are typically developed for multi-access terrestrial settings and do not explicitly address the challenges of TN-NTN backhaul. Focusing on the satellite integration, Shikder et al. [9] proposed a traffic-flow steering algorithm that classifies flows by delay tolerance and redirects suitable traffic to the satellite path under terrestrial congestion. More generally, recent hybrid TNNTN studies emphasize dynamic path selection in response to changing link conditions [10]. However, most of these approaches rely on heuristic rules or focus on generic throughput or congestion objectives without explicitly encoding per-slice SLA penalties. Proactive traffic steering in O-RAN has been shown to reduce queuing delay relative to reactive methods, and networkaided intelligent frameworks optimize flow splitting using utility-based formulations [13]. Traffic classification for traffic steering using Deep Packet Inspection (DPI) or machine learning introduces additional processing overhead at the network edge [14]. Our solution avoids these bottlenecks entirely by performing SLA-aware traffic allocation at the IP transport layer using only aggregated utility metrics. Slice identification is performed by lightweight parsing of the encapsulated packet headers, which is less computationally demanding than full payload DPI. Despite these advances, existing studies lack a unified optimization framework for hybrid TN-NTN backhaul. They neither enforce heterogeneous SLA constraints across slices nor explicitly model the coupled interaction between uplink and downlink traffic competing for shared resources. In contrast, our approach embeds SLA penalties directly into perslice utility functions and solves the problem via a provably

convergent exact potential game, guaranteeing stable, decentralized adaptation under non-stationary loads. III. S YSTEM M ODEL AND G AME T HEORY F RAMEWORK We consider a hybrid TN–NTN backhaul architecture as illustrated in Figure 1. Two independent traffic-steering agents (downlink and uplink) operate at opposite ends of the backhaul. Each agent dynamically allocates traffic fractions across a low-latency terrestrial fiber link and a high-latency, capacityconstrained NTN path. The agents make decisions locally using only their own observations and interface-level telemetry without exchanging any control messages.

Satellite

gNB

NTN Link 5G Core Uplink

Fiber Link Downlink

gNB Loadbalancer (Uplink Agent)

Core Loadbalancer (Downlink Agent)

GTP Tunnel (UDP)

Internet

Fig. 1. System overview of the considered hybrid backhaul, where traffic is dynamically steered between terrestrial fiber and NTN satellite paths.

Fractional traffic splitting is enforced dynamically at the IP layer. The load balancers identify the corresponding slice type of individual UDP-based GPRS Tunneling Protocol (GTP) flows by inspecting packet headers within the encapsulated payload and apply routing rules to distribute traffic across the two backhaul paths. Each slice or UE-group u is assigned a relative priority weight πu , and a set of transport-layer SLA constraints, namely maximum Round-Trip Time (RTT), jitter, and packet loss. While 5G SLAs are traditionally defined end-to-end, we focus exclusively on the backhaul transport layer. The RAN segment is treated as a transparent conduit, and the controller optimizes only routing over the TN and NTN links. A. Game-Theoretic Framework Based on the two-agent definition, we define a noncooperative game as a triple o n   Γ = N , Ai i∈N , Ji i∈N , where N = {DL, UL} is the set of players, Ai is the strategy space of player i, and Ji is the payoff function of player i.

a) Players: The downlink agent controls downstream traffic allocation for the N slices in the set UDL = {1, . . . , N }, and the uplink agent controls upstream allocation for the M slices in the set UUL = {1, . . . , M }. b) Strategies: Each player selects a vector of perslice NTN allocation fractions. For the downlink, αDL = DL (α1DL , . . . , αN ) ∈ [0, 1]N , where αuDL ∈ [0, 1] denotes the fraction of slice u’s downlink traffic routed over the NTN link (and the remainder 1 − αuDL traverses the fiber link). UL Analogously, αUL = (α1UL , . . . , αM ) ∈ [0, 1]M for the uplink agent. The joint strategy space is A = [0, 1]N × [0, 1]M , a compact and convex set. c) Payoff functions: Each player’s payoff function is a weighted sum of per-slice utilities minus a shared capacity coupling penalty. Let αu ∈ {αuDL , αuUL } denote the respective player’s allocation for slice u; its utility Uu (αu ) is defined as cu cu Uu (αu ) = wTcu Tu (αu ) + wL Lu (αu ) + wR Ru (αu )

− wPcu λcu Pu (αu ),

(1)

cu cu , wPcu , wR wTcu , wL cu

≥ 0 are traffic-class-specific where weights, λ ≥ 0 is a class severity multiplier, and the four component terms are defined in Table I. Here Ĉ fib and Ĉ ntn are the adaptive capacity estimates of the fiber and NTN links, respectively; dfib , dntn are the measured round-trip latencies; dmax cu is the SLA latency ceiling for traffic class cu ; ℓfib and ℓntn are the fractional packet loss rates; mk (αu ) = αu metrick,ntn +(1−αu ) metrick,fib is the blended SLA metric on dimension k ∈ {rtt, loss, jitter}; and scku = SLAcku is the corresponding SLA threshold. The capacity coupling penalty C(αDL , αUL ) captures the joint oversubscription cost shared by both agents. For each agent i ∈ {DL, UL}, the load contributions to the NTN and fiber links are X X i i Dntn (αi ) = αui b̂iu , Dfib (αi ) = (1 − αui ) b̂iu , (2) u∈Ui

u∈Ui

with b̂iu

> 0 denoting estimated peak throughput demands. The aggregate link demands are DL UL Dntn (αDL , αUL ) = Dntn (αDL ) + Dntn (αUL ),

Dfib (α

DL

UL

,α

DL UL ) = Dfib (αDL ) + Dfib (αUL ),

(3) (4)

and the coupling penalty is 
2 C(αDL , αUL ) = max 0, Dntn (αDL , αUL ) − Ĉ ntn + 
 2 max 0, Dfib (αDL , αUL ) − Ĉ fib . (5) Let µ > 0 be the coupling coefficient governing the aggressiveness of oversubscription penalization. The payoff functions of the two agents are then X JDL (αDL , αUL ) = πu Uu (αuDL ) − µ C(αDL , αUL ), u∈UDL

(6) JUL (α

DL

UL

,α

)=

X

πv Uv (αvUL ) − µ C(αDL , αUL ).

v∈UUL

(7)

TABLE I C OMPONENT TERMS OF THE PER - SLICE UTILITY Uu (αu ). Term

Formula

Interpretation

Tu (αu )

ln 1 + αu Ĉ ntn + (1 − αu )Ĉ fib

Lu (αu )

−

Ru (αu )

αu (1−ℓntn )+(1−αu )(1−ℓfib )

Reliability (expected packet survival)

!2

Pu (αu )

P

SLA violation penalty (quadratic barrier)

Latency utility (normalised, negative cost)

αu dntn + (1 − αu )dfib dmax cu

max(0, mk (αu ) − scku ) k

Throughput utility (log, proportional fairness)

scku

To achieve a Nash equilibrium, each agent repeatedly solves its best-response problem while the other agent fixes its strategy: αDL∗ = arg αUL∗ = arg

max

JDL (αDL , αUL∗ ),

(8)

max

JUL (αDL∗ , αUL ).

(9)

αDL ∈[0,1]N αUL ∈[0,1]M

Remark III.1 (Decentralized execution). Since backhaul traffic is bidirectional, each agent’s RX counters on the NTN and fiber interfaces directly measure the opposing agent’s aggregate load. Each agent therefore solves (8)–(9) using only local telemetry, without inter-agent communication. B. Exact Potential Game Formulation We now formalize the game as an exact potential game by constructing an explicit potential function and proving the alignment condition of Monderer and Shapley [15]. Definition III.2 (Nash equilibrium). A joint strategy profile α∗ = (αDL∗ , αUL∗ ) ∈ [0, 1]N ×[0, 1]M is a Nash equilibrium of the game Γ if and only if JDL (αDL∗ , αUL∗ ) ≥ JDL (αDL , αUL∗ ),

∀ αDL ∈ [0, 1]N , (10)

JUL (αDL∗ , αUL∗ ) ≥ JUL (αDL∗ , αUL ),

∀ αUL ∈ [0, 1]M . (11)

At a Nash equilibrium, neither agent can unilaterally improve its own payoff by deviating from α∗ . Definition III.3 (Exact Potential Game [15]). A game Γ = {N , {Ai }, {Ji }} is an exact potential game if there exists a function Φ : A → R, called an exact potential function, such that for every player i ∈ N and for every two strategy profiles that differ only in player i’s action: Ji (ai , a−i )−Ji (a′i , a−i ) = Φ(ai , a−i )−Φ(a′i , a−i ), (12) where a−i is the strategies of all players except player i. The key implication of an exact potential game is that any unilateral improvement by one agent increases Φ by the exact same amount as it increases that agent’s own payoff. As a

result, best-response dynamics steadily ascend the potential function, ensuring convergence to a pure Nash equilibrium. Theorem III.4 (Exact Potential Function). Let α = (αDL , αUL ) denote the joint allocation vector. Given the payoff functions JDL and JUL , the load-balancing game Γ is an exact potential game with an exact potential function

The diagonal matrix D(α) has entries, for each slice u, " wTcu (∆Ĉ)2 Duu (α) = −πu
2 + 1 + Ĉ fib + αu ∆Ĉ # X 2(∆mk )2 cu cu cu 1[mk (αu ) > sk ] ≤ 0, (16) wP λ (scku )2 k

Φ(α) =

X

πu Uu (αuDL ) +

X

πv Uv (αvUL ) − µ C(α).

v∈UUL

u∈UDL

(13) Proof. It can be verified by direct substitution that for any unilateral deviation of one agent (holding the other fixed), the change in the payoff of that agent equals the change in Φ, satisfying the exact potential condition of [15]. C. Equilibrium Analysis and Convergence Having established that the load-balancing game is an exact potential game, we now prove that it admits a unique Nash equilibrium by leveraging the strict concavity of the potential function Φ.

where ∆mk = metrick,ntn − metrick,fib and the inequality holds because wTcu , wPcu , λcu , πu ≥ 0. Under the standing assumption that Ĉ ntn ̸= Ĉ fib (NTN and fiber have different capacities), we have ∆Ĉ ̸= 0, and therefore Duu (α)

≤

Lemma III.6 (Uniqueness of Nash equilibrium). The game Γ has a unique Nash equilibrium α∗ ∈ [0, 1]N +M . Proof. We verify both conditions as included in [15]: Condition (i): Compact and convex strategy space. Each per-slice allocation αu ∈ [0, 1] is confined to a closed, bounded interval. Thus, A = [0, 1]N × [0, 1]M = [0, 1]N +M is a compact convex subset of RN +M . Condition (ii): Continuous differentiability and strict concavity of Φ. The potential Φ is continuously differentiable on the interior (0, 1)N +M as every component (Tu is a log of an affine function, Lu and Ru are affine, Pu and the coupling term C(α) are sums of squared-hinge functions) is C 1 . For strict concavity, we compute the Hessian HΦ (α) ∈ R(N +M )×(N +M ) and show it is negative definite. Since each per-slice utility Uu (αuDL ) depends only on the scalar αuDL and analogously for the UL, the Hessian of the aggregate utility terms is diagonal. Accordingly, HΦ (α) = D(α) − µ κ(α) b+ (b+ )⊤ , | {z }

(14)

diagonal

DL UL UL ⊤ > 0 is the where b+ = (b̂DL 1 , . . . , b̂N , b̂1 , . . . , b̂M ) vector of estimated peak demands, and κ(α) ∈ {0, 2, 4} is a non-negative scalar arising from the activation of the oversubscription indicators:     κ(α) = 2(1 Dntn > Ĉ ntn + 1 Dfib > Ĉ fib ). (15)

wTcu (∆Ĉ)2 1 + Ĉ fib + αu ∆Ĉ

2

<

0.

(17)

Hence D(α) ≺ 0 (strictly negative definite diagonal matrix). Furthermore, −µκ(α) b+ (b+ )⊤ is a negative semidefinite rank-1 correction. Therefore HΦ (α) ≺ 0 for all α ∈ (0, 1)N +M , and since Φ is continuous on the compact set [0, 1]N +M , it follows that Φ is strictly concave on [0, 1]N +M . Since both conditions hold, the Nash equilibrium of the load-balancing game Γ is unique.

Lemma III.5 (Existence of Nash equilibrium). A purestrategy Nash equilibrium exists for the game Γ. Proof. Since Γ is an exact potential game and the strategy space A = [0, 1]N +M is compact, a maximizer of Φ over A exists by the Extreme Value Theorem. Every maximizer of an exact potential function is a Nash equilibrium [15].

−πu ·

IV. E XPERIMENTAL S ETUP The experimental evaluation is conducted on a geographically distributed 5G testbed, with the RAN deployed in Malaga and the core deployed in Ghent. This setup emulates a hybrid backhaul scenario in which TN and NTN links coexist and are jointly used for traffic delivery. As depicted in Figure 1, a set of UE devices connect to Nokia Airscale gNB, which forwards uplink traffic to an edge-based load-balancing virtual machine acting as the uplink routing agent. This edge agent connects to a corresponding downlink load balancer at the core via two parallel virtual tunnels: a low-latency terrestrial fiber link and an NTN satellite link utilizing Starlink. To evaluate the framework defined in Section III, bidirectional traffic is generated to emulate five distinct slices: V2X, emergency services, video streaming, IoT, and best-effort web traffic. We configured the testbed with the specific transportlayer SLA constraints detailed in Table II. These values are configured to reflect the geographic distance and the physical characteristics of the Starlink NTN environment. Generated flows representing the five slices are explicitly non-stationary, incorporating randomized burst durations, pause intervals, and variable bandwidth demands. TABLE II SLA CONSTRAINTS PER TRAFFIC CLASS . Traffic Class

Priority

Max RTT (ms)

Max Jitter (ms)

Max Loss (%)

V2X Emergency Video Streaming IoT Best-effort

High High Medium Low Low

60 70 200 500 800

15 20 80 150 100

0.5 0.5 3.0 10.0 5.0

Table III summarizes the average network metrics recorded across the 300-second non-stationary traffic episodes. Here, the RTT represents the effective RTT proportionally weighted by the actual volume of traffic steered over the TN and NTN links. Similarly, the reported throughput reflects the total bidirectional traffic achieved by the system under the same traffic scenario. The table shows that the potential game framework consistently outperforms the other algorithms in maximizing overall network efficiency. It achieved the lowest effective RTT (51.683 ms), the highest bidirectional throughput (78.195 Mbps), and the lowest packet loss rate of 0.2%. Although these aggregate improvements are modest, the primary benefit lies in the substantial reduction in SLA violations per slice, as shown separately in the heatmap of Figure 2. Nonetheless, this marginal improvement highlights the benefits of the utility evaluation to proactively avoid congestion. Fairness is calculated using Jain’s fairness index [16], applied to per-link throughput utilization across the two backhaul paths. The Equal Split baseline naturally achieved a perfect fairness score of 1.000, as it indiscriminately distributes traffic. However, enforcing strict arithmetic fairness across highly asymmetric links increases packet loss and RTT. By forcing 50% of latency-sensitive and high-bandwidth traffic onto the constrained satellite link, Equal Split yields the highest packet loss (1.445%) among all methods. In contrast, the potential game achieves a fairness score of 0.963 while simultaneously delivering the best throughput, latency, and reliability, proving that the algorithm successfully balances individual slice requirements without starving any specific flow. While aggregate metrics provide a macro-level view of system efficiency, the controller’s primary objective is to

Algorithm

RTT (ms)

Loss (%)

Throughput (Mbps)

Fairness

Potential Game 51.683 Heuristic 53.884 Equal Split 54.801 Weighted RR 53.463 Random 54.793

0.200 0.349 1.445 0.901 1.164

78.195 75.307 75.193 74.374 70.002

0.963 0.855 1.000 0.709 0.732

maintain strict SLA compliance across heterogeneous slices. Figure 2 displays a heatmap of the SLA violation for each load-balancing algorithm over the evaluation period. The potential game controller demonstrates superior performance in safeguarding strict SLA requirements. For lesssensitive slices, such as IoT, video, and best effort, it completely eliminates SLA violations. For sensitive slices, it reduces the SLA violations to 1.7% for V2X and 0.7% for the emergency slice. These residual violations occur only during severe congestion spikes, when even the optimal allocation cannot fully prevent brief queuing delays on the terrestrial link for slices with tight SLA requirements. The IoT slice remains highly robust across almost all configurations. This indicates that its loose SLA requirements are easily satisfied across the hybrid backhaul. The primary challenge in backhaul load-balancing is intelligently steering ultra-low-latency V2X and emergency flows alongside highbandwidth video and best-effort flows to avoid the NTN latency bottleneck and terrestrial link congestion. Finally, the temporal evolution of the link-specific metrics and the objective function is depicted in Figure 3. Throughout the evaluation period, the potential function Φ exhibits a stable baseline convergence plateau, punctuated by transient, distinct drops. These drops correlate with mathematical penalties incurred when SLA constraints are breached, primarily driven by RTT degradation (second panel) and packet loss events (third panel) during severe congestion spikes.

SLA Violation Rate [%] per Slice Potential Game

1.7

0.7

0.0

0.0

0.0

Heuristic

9.8

4.7

4.3

0.0

3.2

Equal Split

67.0

62.6

34.0

0.0

3.1

Weighted RR

34.1

45.7

16.9

0.0

1.5

Random

36.0

39.3

13.1

0.7

2.6

y o genc Vide Emer

IoT

Effo Best

60 50

SLA violation rate [%]

V. R ESULTS AND D ISCUSSION

TABLE III AVERAGE NETWORK - LEVEL PERFORMANCE METRICS PER ALGORITHM .

Algorithm

The real-time network state is captured by a custom metrics exporter running natively on the load balancers that exposes continuous telemetry such as interface-level throughput, RTT, jitter, and packet loss to a time-series database. Routing decisions are governed by two separate controllers deployed at the uplink and downlink load balancers. These controllers utilize the Best Response Iteration (BRI) algorithm to solve the proposed potential game and determine the optimal traffic allocation vectors for each slice. Since the potential function Φ is strictly concave, BRI is guaranteed to converge to the unique pure Nash equilibrium. The proposed potential game approach is benchmarked in the testbed against a set of representative baseline routing strategies commonly used in hybrid and slice-aware traffic steering, including equal split, weighted round robin, random allocation, and a purpose-built SLA-aware heuristic. The heuristic baseline computes a per-slice score for each link by combining normalized RTT, loss, and jitter headroom with traffic-class weights on both links and performs redistribution sweeps that shift load from any link exceeding 90% utilization to the alternate link. The main evaluation metrics are aggregate network metrics and per-slice SLA violation rates.

40

V2X

Slice

30 20 10

rt

0

Fig. 2. Heatmap of the average SLA violation rates per traffic slice.

Potential Function

1.6 1.5 1.4 1.3 1.2

Round-Trip Time

RTT [ms]

80

Fiber NTN

70

R EFERENCES

60 50

Packet Loss

Packet Loss [%]

8

Fiber NTN

6 4 2 0

Throughput [Mbps]

Bidirectional Throughput 60 40 20 0

we introduced a decentralized load-balancing controller that models per-slice traffic steering as an exact potential game. Testbed evaluations demonstrated that our approach significantly outperforms heuristic baselines under non-stationary loads, ensuring strict SLA compliance. Future work will adapt this framework for moving LEO constellation dynamics and integrate it with 5G core orchestration for autonomous, crosslayer SLA management.

Fiber NTN 0

50

100

150

Relative time [s]

200

250

300

Fig. 3. Temporal evolution of the potential function and network metrics.

It is critical to note that high latency does not universally penalize the system. Drops in Φ are triggered only when physical network impairments violate an assigned SLA, such as when congestion forces time-sensitive V2X or emergency slices to experience queuing delays or packet drops. Furthermore, Figure 3 highlights the reactive agility of the potential game framework. Upon detecting a sub-optimal utility state (a drop in Φ), the decentralized controller immediately executes new best-response sweeps and shifts the fractional allocation to alleviate the bottleneck. Consequently, the potential function sharply recovers to its maximized equilibrium state. This pattern confirms that the controller continuously tracks a shifting pure Nash equilibrium in real time. The proposed controller exhibits low computational overhead, making it suitable for real-time deployment. Each agent solves a continuous optimization problem over its per-slice allocation variables, with linear complexity O(N ) or O(M ) per iteration. In our testbed deployment, each best-response computation completes in an average of 4.412 ms on a standard desktop CPU (Intel Core i9-9900K). The alternating best-response dynamics typically reach equilibrium in less than five iterations under typical load conditions. VI. C ONCLUSION Integrating hybrid TN-NTN backhauls enhances 5G/6G resilience but complicates SLA management. To address this,

[1] M. M. Azari, S. Solanki, S. Chatzinotas, O. Kodheli, H. Sallouha, A. Colpaert, J. F. Mendoza Montoya, S. Pollin, A. Haqiqatnejad, A. Mostaani, E. Lagunas, and B. Ottersten, “Evolution of Non-Terrestrial Networks From 5G to 6G: A Survey,” IEEE Communications Surveys & Tutorials, vol. 24, no. 4, pp. 2633–2672, 2022. [2] 3rd Generation Partnership Project (3GPP), “Service requirements for the 5G system; Stage 1 (Release 20),” 3GPP Technical Specification Group Services and System Aspects, Technical Specification TS 22.261, March 2026. [3] O. Shan Wong, M. A. Gregory, and S. Li, “Integration of non-terrestrial network for 5G NR and future 6G: LEO satellite-to-device performance and interference analysis,” Computer Networks, vol. 275, p. 111870, 2026. [4] B. Di, H. Zhang, L. Song, Y. Li, and G. Y. Li, “Ultra-Dense LEO: Integrating Terrestrial-Satellite Networks Into 5G and Beyond for Data Offloading,” IEEE Transactions on Wireless Communications, vol. 18, no. 1, pp. 47–62, 2019. [5] H. Zhang, N. Liu, X. Chu, K. Long, A.-H. Aghvami, and V. C. M. Leung, “Network Slicing Based 5G and Future Mobile Networks: Mobility, Resource Management, and Challenges,” IEEE Communications Magazine, vol. 55, no. 8, pp. 138–145, 2017. [6] D. Pugliese, M. Quadrini, D. Striccoli, C. Roseti, F. Zampognaro, G. Piro, L. A. Grieco, and G. Boggia, “Integrating terrestrial and non-terrestrial networks via IAB technology: System-level design and evaluation,” Computer Networks, vol. 253, p. 110726, 2024. [7] H. Xu, Z. Shi, M. Liu, N. Zhang, Y. Yan, and G. Han, “Link-State Aware Hybrid Routing in the Terrestrial–Satellite Integrated Network,” Sensors, vol. 22, no. 23, p. 9124, 2022. [8] I. Sawad, R. Nilavalan, and H. Al-Raweshidy, “Backhaul in 5G systems for developing countries: A literature review,” IET Communications, vol. 17, no. 6, pp. 659–669, 2023. [9] M. S. Shikder, M. M. Hasan, S. H. Wakib, A. Tasnim, and M. S. Rahman, “A Traffic Flow Steering Algorithm for Hybrid TerrestrialSatellite Backhaul Network,” in Proceedings of the 2nd International Conference on Computing Advancements, ser. ICCA ’22. Association for Computing Machinery, 2022, pp. 187–190. [10] S. B. R. Tirmizi, Y. Chen, S. Lakshminarayana, W. Feng, and A. A. Khuwaja, “Hybrid Satellite–Terrestrial Networks toward 6G: Key Technologies and Open Issues,” Sensors, vol. 22, no. 21, p. 8544, 2022. [11] Y. T. Seyoum, S. M. Shahid, E. S. Cho, and S. Kwon, “Distributed load balancing algorithm considering QoS for next generation multiRAT HetNets,” Computer Networks, vol. 229, p. 109758, 2023. [12] X. Ba, L. Jin, Z. Li, J. Du, and S. Li, “Multiservice-Based Traffic Scheduling for 5G Access Traffic Steering, Switching and Splitting,” Sensors, vol. 22, no. 9, p. 3285, 2022. [13] I. Tamim, S. Aleyadeh, and A. Shami, “Intelligent O-RAN Traffic Steering for URLLC Through Deep Reinforcement Learning,” 2023. [Online]. Available: https://arxiv.org/abs/2303.01960 [14] A. Azab, M. Khasawneh, S. Alrabaee, K.-K. R. Choo, and M. Sarsour, “Network traffic classification: Techniques, datasets, and challenges,” Digital Communications and Networks, vol. 10, no. 3, pp. 676–692, 2024. [15] D. Monderer and L. S. Shapley, “Potential Games,” Games and Economic Behavior, vol. 14, no. 1, pp. 124–143, 1996. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0899825696900445 [16] R. Jain, D. Chiu, and W. Hawe. (1998) A Quantitative Measure Of Fairness And Discrimination For Resource Allocation In Shared Computer Systems.