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

TL;DR

The paper addresses how queue-state disclosures influence impatient tenants in decentralized edge and NTN systems and proposes an information-bulletin framework with two Markov descriptors: a service-rate distribution and an Inter-Change Time Distribution, enabling tenants to decide reneging or jockeying. Tenants’ actions feed back to queues, which adapt the service-rate vector $\bar{\mu}$ via a rule-based predictive policy, and the approach is benchmarked against a centralized hedging-point MDP for M/M/2 systems; results emphasize robustness to partial or stale information. The contributions include formalizing the two Markov descriptors, deriving reneging and jockeying probabilities, and demonstrating reduced impatience under information-driven policies, particularly when information is stale. The findings highlight practical potential for robust, decentralized resource management in edge-cloud and NTN deployments, with policy-driven adaptability complementing centralized optimization.

Abstract

We study how queue-state information disclosures affect impatient tenants in multi-tenant edge systems. We propose an information-bulletin strategy in which each queue periodically broadcasts two Markov models. One is a model of steady-state service-rate behavior and the other a model of the queue length inter-change times. Tenants autonomously decide to renege or jockey based on this information. The queues observe tenant responses and adapt service rates via a learned, rule-based predictive policy designed for decentralized, partially-observed, and time-varying environments. We compare this decentralized, information-driven policy to the classical, centralized Markov Decision Process (MDP) hedging-point policy for M/M/2 systems. Numerical experiments quantify the tradeoffs in average delay, impatience and robustness to stale information. Results show that when full, instantaneous state information and stationarity hold, the hedging-point policy yields less impatience but this diminishes as information becomes partial or stale. The rule-based predictive policy on the other hand is more robust to staleness in dispatched information, making it conducive for conditions typical of edge cloud and non-terrestrial deployments.

Figures (5)

• Figure 1: When no policy is embedded, the resulting phenomena here is instability regardless of the Markov model of information dispatched. This volatility however as expected reduces when the dispatch intervals are increased.
• Figure 2: The rule-based queue policy that learns the dispatching and service rates manages to regulate the impatience with the Markov model of the service rates proving more vital especially in keeping the jockeying minimal.
• Figure 3: Across all intervals, slightly lower and more stable renege rates are observed for both policies, though the hedge-point policy remains relatively higher. The jockeying rates for all policies begin to converge at higher dispatch intervals and show less volatility.
• Figure 4: The box plot showing the median waiting time when a policy is embedded versus otherwise for both Markov information models. The waiting time and spread here are high when no policy is embedded for both impatience behavior kiggundu2025information.
• Figure 5: Here, the rule-based policy leads to substantially better (lower) objective values in comparison to when no policy is embedded. Our policy finds the near-optimum despite changes in the interval and impatience. Absence of a control policy on the other hand is consistently suboptimal under the behavioral changes since it cannot adapt.