Adaptive Decentralized Queue Disclosure for Impatient Tenants in Edge and Non-terrestrial Systems

Basic Information

• Paper ID: 2508.04241
• Title: Adaptive Decentralized Queue Disclosure for Impatient Tenants in Edge and Non-terrestrial Systems
• Authors: Anthony Kiggundu, Bin Han, Hans D. Schotten
• Classification: eess.SY (Systems and Control), cs.SY (Systems and Control)
• Publication Date: October 13, 2025 (arXiv v2)
• Institutions: German Research Center for Artificial Intelligence (DFKI), RPTU University of Kaiserslautern-Landau
• Paper Link: https://arxiv.org/abs/2508.04241

Abstract

This paper investigates how queue state information disclosure affects impatient tenants in multi-tenant edge systems. The authors propose an information announcement strategy where each queue periodically broadcasts two Markov models: one characterizing steady-state service rate behavior and another modeling queue length change timing. Tenants autonomously decide whether to abandon or switch queues based on this information. Queues observe tenant responses and adapt service rates through a learning-based rule-driven prediction strategy designed for decentralized, partially observable, and time-varying environments. Numerical experiments quantify trade-offs between average delay, impatience levels, and robustness to stale information.

Research Background and Motivation

Problem Definition

In heterogeneous 5G/6G deployments, multi-tenant resource sharing is driven not only by static configurations but increasingly by autonomous tenant decisions (e.g., whether to offload tasks to remote queues or process locally). Queue state disclosure (such as queue length, waiting time estimates, or service statistics) can significantly alter tenant behavior and trigger resource competition through queue jockeying and reneging.

Research Significance

Modern multi-access edge computing (MEC) and non-terrestrial network (NTN) environments are decentralized, characterized by partial and stale state broadcasts, and exhibit time-varying channels and mobility. In such environments, assuming a single central controller with instantaneous global state is unrealistic. However, existing disclosure rules and heuristics are typically developed for static or lightly mobile settings and fail to address three fundamental questions in decentralized control:

1. What state information should be shared
2. How should information be represented
3. How frequently should updates be distributed

Limitations of Existing Approaches

Traditional centralized optimization methods (such as hedging point policies) assume complete, instantaneous state information and stationarity conditions that often do not hold under typical edge cloud and non-terrestrial deployment conditions. Existing approaches show significantly degraded performance when information becomes partial or stale.

Core Contributions

1. Information Announcement Concept: Introduces the information announcement concept for multi-tenant queues and formalizes two Markov descriptors (service rate distribution and change timing) as tunable state summaries suitable for resource-constrained control channels.
2. Theoretical Analysis: Derives closed-form expressions for queue jockeying and reneging probabilities under these descriptors and formulates a joint impatience minimization problem balancing delay, jockeying, and reneging. Proves the optimization problem is analytically intractable.
3. Practical Strategy: Proposes a practical rule-based prediction strategy that learns service rate vectors from tenant responses and adapts service rates online.
4. Comprehensive Evaluation: Provides extensive numerical evaluation quantifying the value of different announcement models and distribution intervals, demonstrating the robustness of the learning strategy under heterogeneous workloads.

Methodology Details

Task Definition

Consider an M/M/2 queueing system with two queues i and j. New arrivals follow a Poisson distribution with total arrival rate λ = λᵢ + λⱼ. Each queue distributes its state information to tenants at interval r seconds, introducing staleness. The objective is to minimize a composite performance measure of average delay, jockeying events, and reneging (tenant impatience).

Model Architecture

1. Markov Service Rate Model

The service rate distribution of queue i or j in steady state follows a K-state continuous-time Markov chain (CTMC) with service rates {μᵢ}ᵢ₌₁ᴷ and {μⱼ}ⱼ₌₁ᴷ. The effective service rate is defined as:

μ̄ₓ = Σᵢ₌₁ᴷ πₓᵢ μᵢ, μ̄ᵧ = Σⱼ₌₁ᴷ πᵧⱼ μⱼ

where πₓᵢ and πᵧⱼ are steady-state probabilities.

2. Queue Length Dynamics Model - Inter-Change Distribution (ICD)

This model quantifies the frequency of transitions in the queue system. For a queue in state n, only arrival events change state when n=0, while both arrival and departure events may occur when n≥1. The Markov model is defined as:

Rᵢ = Σₙ₌₀^∞ πᵢ,ₙ (λᵢ + μᵢ · 1ₙ≥₁) = 2λᵢ

The expected inter-change interval time is:

Tᵢᴵᶜᴰ = 1/Rᵢ = 1/(2λᵢ)

3. First-Order Stochastic Dominance (FSD)

Better queues are determined by comparing cumulative distribution functions FX(μₖ) and FY(μₖ). X first-order stochastically dominates Y if PX > x ≥ PY > x ∀x ∈ ℝ.

Behavior Modeling

Reneging Behavior

The reneging probability based on FSD is defined as:

P^FSD_reneg(ℓ) = Σᵥ₌₀^(ℓ-1) [(μᵢ - λᵢ)Δ]^v/v! e^(-(μᵢ-λᵢ)Δ)

where Δ = Tₗₒcₐₗ - ηr, and η ∈ 0,1 represents the degree of information staleness.

Jockeying Behavior

The jockeying probability based on ICD is modeled using a sigmoid function:

P^ICD_{i→j} = 1/(1 + e^(-2de^(-ηr)(λᵢ-λⱼ)))

Optimization Problem

The joint optimization problem is formalized as:

min_{μᵢ,μⱼ} τ[Wᵢ(μᵢ) + Wⱼ(μⱼ)] + φ[R^reneg_i(μᵢ) + R^reneg_j(μⱼ)] + ψ[R^jockey_{i→j}(μᵢ,μⱼ) + R^jockey_{j→i}(μⱼ,μᵢ)]

Subject to constraints: μᵢ,min ≤ μᵢ < μᵢ,max, μᵢ > λᵢ

Technical Innovations

1. Information Abstraction: Abstracts complex queue states into two compact Markov models suitable for bandwidth-limited control channels.
2. Adaptive Learning: The rule-based prediction strategy learns from tenant responses and adapts service rates online.
3. Robustness Design: Accounts for information staleness and partial observability, better suited for practical edge computing environments.

Experimental Setup

Experimental Parameters

• Distribution intervals: r ∈ {3, 5, 7, 9} seconds
• Arrival rate range: 3 ≤ λ ≤ 17
• 300 simulation runs per configuration
• M/M/2 system setup

Evaluation Metrics

• Average delay
• Reneging rate
• Jockeying rate
• Composite objective function value (combining delay and impatience measures)

Comparison Methods

• No-strategy baseline
• Classical centralized MDP hedging point policy
• Proposed rule-based prediction strategy

Experimental Results

Main Findings

1. Information Model Comparison: The Markov service rate model produces less impatient behavior than the queue length change timing model, as it provides direct mapping of processing speed.
2. Distribution Frequency Optimization: Optimal performance is achieved at 5-7 second intervals, where impatience is minimized and system stability is maintained, particularly when requests obtain service rate information.
3. Strategy Comparison:
   • Hedging point policy: More stable but higher reneging and jockeying rates
   • Rule-based strategy: More variable but may record lower rates at smaller intervals
4. Optimization Effectiveness: The optimized strategy is statistically robust, producing lower and more consistent objective values (mean=0.53 vs 1.78 without optimization).

Key Findings

According to the quantitative summary in Table I:

• Optimized results show lower variability (standard deviation=0.15 vs 0.97)
• Average improvement of 1.26
• Better solutions found across all distribution intervals

Waiting Time Analysis

When strategies are embedded, waiting times for reneging and jockeying requests are significantly reduced, with greater optimality observed particularly when distributing the Markov service rate model.

Related Work

Major research directions in this field include:

1. Information disclosure strategies in queueing systems
2. Decentralized control for multi-server systems
3. Resource allocation in edge computing
4. Behavior modeling of impatient customers

Advantages of this paper over related work:

• Accounts for the impact of information staleness
• Provides solutions suitable for decentralized environments
• Incorporates learning and adaptation mechanisms

Conclusions and Discussion

Main Conclusions

1. System state information plays a critical role in shaping impatient tenant decisions
2. Rule-based prediction strategies demonstrate greater robustness to information staleness
3. Appropriate information disclosure frequency is crucial for system performance
4. Markov service rate models are more effective than queue dynamics models

Limitations

1. Limited to M/M/2 Poisson settings
2. Requires quantification of computational and communication overhead of announcement mechanisms
3. Does not consider bursty, heavy-tailed arrival processes and non-exponential service times

Future Directions

1. Incorporate information models with explicit announcement subscription costs
2. Replace rule-based heuristics with reinforcement learning techniques
3. Extend to multi-queue heterogeneous servers
4. Validate methods on prototype MEC testbeds

In-Depth Evaluation

Strengths

1. Novelty: Proposes a novel information announcement concept, offering new perspectives for decentralized queue control
2. Practicality: Accounts for information staleness and partial observability in real edge computing environments
3. Theoretical Rigor: Provides a complete mathematical modeling and analysis framework
4. Comprehensive Experiments: Validates method effectiveness through extensive numerical experiments

Weaknesses

1. Model Limitations: Considers only M/M/2 systems; real-world systems are more complex
2. Parameter Sensitivity: Selection of certain parameters (e.g., δλ, η) lacks sufficient theoretical guidance
3. Computational Complexity: Insufficient detail on computational complexity analysis of KKT condition solving
4. Practical Validation: Lacks validation experiments on real systems

Impact

1. Academic Contribution: Provides new research directions for queueing theory and edge computing
2. Practical Value: Offers guidance for resource allocation in 6G networks
3. Scalability: The methodological framework demonstrates good extensibility

Applicable Scenarios

This method is particularly suitable for:

1. Multi-tenant edge computing systems
2. Non-terrestrial network environments
3. Decentralized systems with limited information transmission
4. Service systems requiring consideration of user impatience behavior

References

The paper cites important literature from queueing theory, behavior modeling, and edge computing, including:

• Y. Ouyang and D. Teneketzis's research on decentralized routing signaling
• B. Lin et al.'s work on optimal policies for dual-server queueing systems
• 3GPP technical specifications on network slicing management and orchestration

Overall Assessment: This is a high-quality research paper at the intersection of queueing theory and edge computing, proposing innovative information disclosure strategies to address tenant impatience in decentralized environments. Despite certain limitations, its theoretical contributions and practical value make it an important advance in the field.