Data-Driven Queue Management Mechanisms

• Data-driven queue management is a design principle that transforms queue states into structured records and embeddings to drive admission, ordering, and control actions.
• It integrates fixed-rule heuristics, model-driven optimization, and learning-based control to adaptively manage queues in response to real-time conditions.
• Empirical studies in domains from embedded systems to packet networks demonstrate improvements in throughput, latency, and fairness, guiding future hybrid designs.

Searching arXiv for the provided and closely related queue-management papers to ground the article with current citations. arxiv_search.query={"^^^^^^^^4search_query4 queue management\" OR ^^^^^^^^4all:\4 queue management\" OR ^^^^^^^^4all:\4 management mechanism\"","start":^^^^4search_query4^^^^,"max_results":^^^^4all:\4search_query4 Data-driven queue management denotes a family of mechanisms in which queue admission, ordering, service, signaling, or routing decisions are derived from structured observations about queued entities and service conditions, rather than from unstructured token passing or fixed rules alone. In the cited literature, the term spans low-cost embedded patient intake backed by structured records, delay- and backlog-aware Active Queue Management (AQM) in packet networks, vision-based queue-state estimation in public spaces, demand-aware task release in multi-tenant clusters, Age-of-Information (AoI)-aware packet dropping in vehicular links, and diffusion-approximation-based skill routing in large call centers (&&&^^^^4search_query4^^^^&&&, &&&^^^^4all:\4^^^^&&&, &&&^^^^4 OR all:\4^^^^&&&, &&&^^^^4 OR all:\4^^^^&&&, Fang et al., 27 Jan 2026, Ata et al., 10 May 2026). The unifying idea is explicit state representation: queue positions become records, vectors, or learned embeddings that can drive enqueue, dequeue, drop, mark, or matching decisions.

^^^^4all:\4^^^^. Conceptual scope

The literature does not use “data-driven” in a single sense. Some systems are data-driven because each queue entry is backed by a structured record. In the embedded medical-chamber system, the queue is a circular buffer of patient records containing serial number, demographics, temperature, and pulse/BPM, and the doctor retrieves the next record with a single button press (&&&^^^^4search_query4^^^^&&&). Other systems are data-driven in the classical feedback-control sense: decisions depend on measured queue occupancy, average queue size, rate of change, delay, or RTT rather than on static thresholds alone. Examples include the HSDPA Time Space Priority mechanism with state-dependent non-real-time admission, AQMRD’s use of both average queue size and its rate of change, and FuzzyRTT’s RTT-aware fuzzy control (Hanini et al., 2013, Karmeshu et al., 2016, Boudi et al., 2024).

A third meaning is model-driven optimization. PAQMAN formulates admit/drop control as a Semi-Markov Decision Process calibrated from observed arrivals and RTT-related conditions, while the high-dimensional call-center scheduler calibrates a many-class queueing model from operational data and solves a diffusion-control approximation via deep neural networks (&&&^^^^4all:\4search_query4^^^^&&&, Ata et al., 10 May 2026). A fourth meaning is explicitly learning-based control: SAM uses an SVM over queue-size patterns, the ECN-based intelligent AQM couples LSTM prediction with Q-learning, AQM-LLM distills a pretrained LLM into an L4S controller, and the V^^^^4 OR all:\4^^^^V AoI framework learns joint dropping and power-control actions with GNN-enhanced MAPPO (&&&^^^^4all:\4 OR all:\4^^^^&&&, &&&^^^^4all:\4 OR all:\4^^^^&&&, &&&^^^^4all:\44^^^^&&&, Fang et al., 27 Jan 2026).

Taken together, these papers suggest a continuum rather than a dichotomy. At one end are record-backed FIFO systems whose service order remains fixed; at the other are learned controllers whose actions alter queue evolution itself. A common misconception is therefore that “data-driven” necessarily means adaptive reprioritization. Several systems are data-rich yet intentionally non-adaptive in service order, because the data support identity, visibility, or consultation context rather than dynamic scheduling (&&&^^^^4search_query4^^^^&&&).

^^^^4 OR all:\4^^^^. Observed state and data structures

Across domains, the queue state is rarely a scalar. In the embedded patient system, each record contains patient ID or serial number, age, encoded name or short code, mobile number, temperature, and pulse/BPM, all held in volatile SRAM in a circular buffer (&&&^^^^4search_query4^^^^&&&). In Tromino, the relevant queue-side state is not merely task count but aggregate resource demand, expressed through Dominant Share (current consumption) and Dominant Demand Share (queued demand) per framework (&&&^^^^4 OR all:\4^^^^&&&). In the call-center scheduler, the observable service-system state includes class populations PRESERVED_PLACEHOLDER_^^^^4search_query4^^^^, waiting counts PRESERVED_PLACEHOLDER_^^^^4all:\4^^^^, idleness PRESERVED_PLACEHOLDER_^^^^4 OR all:\4^^^^, and activity allocations PRESERVED_PLACEHOLDER_^^^^4 OR all:\4^^^^, which are then centered and diffusion-scaled around a static planning solution (Ata et al., 10 May 2026).

Packet-network proposals enrich queue state even further. “Rapid Signalling of Queue Dynamics” requires enqueue timestamp, dequeue timestamp, backlog at enqueue, and backlog at dequeue to compute a fresher queue-time signal (&&&^^^^4all:\4^^^^&&&). The ECN-based intelligent AQM uses counts of ECE-marked packets over fixed intervals, with a ten-step history as the input to an LSTM congestion predictor (&&&^^^^4all:\4 OR all:\4^^^^&&&). AQM-LLM defines an 8-feature state comprising queue type, burst allowance, drop probability, current queue delay, accumulated probability, total bytes, total drops, and packet length (&&&^^^^4all:\44^^^^&&&). PAQMAN, by contrast, abstracts the queue into queue length and inferred arrival-rate parameters, with RTT added in the non-negligible-delay formulation (&&&^^^^4all:\4search_query4^^^^&&&).

In public-space and cyber-physical queueing, state can be geometric or freshness-oriented rather than purely occupancy-based. Massimo extracts hip midpoints from YOLO pose estimation, fits an “optimal line,” and computes regression residuals, pairwise distances, angles, and force magnitudes from a mass-spring model (&&&^^^^4 OR all:\4^^^^&&&). The V^^^^4 OR all:\4^^^^V AoI framework represents each transmitter queue by at most two interdependent batches, with remaining packet counts $q_m^{(1)}[n]$, $q_m^{(2)}[n]$, transmitter-side ages, receiver AoI, and a GNN embedding of the slowly varying interference topology (Fang et al., 27 Jan 2026).

Domain	State representation	Decision target
Embedded medical chamber	Patient record: serial, demographics, temperature, pulse	FIFO enqueue/dequeue
Packet AQM	Backlog, delay, RTT, ECN, drop/mark history	Drop, mark, or threshold choice
Public queue monitoring	Hip midpoints, line fit, force scores	Outlier/crowding detection
Cluster scheduling	DS, DDS, queued task demand	Task release by framework
AoI-sensitive wireless	Batch counts, ages, channel state, topology embedding	Drop/keep and power control
Parallel-server services	$X$, $Y$, $Z$, $\psi$, diffusion state	Skill-based routing

This breadth matters because it shows that “queue data” may encode identity, congestion, geometry, fairness, or freshness, depending on the operational objective. The design of a data-driven queue manager is therefore inseparable from the choice of state variables that are considered decision-relevant.

^^^^4 OR all:\4^^^^. Control laws and decision paradigms

The control laws reported in the literature range from explicit queue equations to implicit learned policies. Some are deliberately simple. The ATmega^^^^4 OR all:\4 OR all:\4^^^^ patient system enforces strict first-come, first-served ordering with enqueue after successful registration and dequeue on doctor button press; the physiological measurements are recorded and displayed but do not influence service order (&&&^^^^4search_query4^^^^&&&). Tromino similarly preserves a simple release logic at the framework boundary, but changes which framework is allowed to release work based on DS, DDS, and current cluster availability (&&&^^^^4 OR all:\4^^^^&&&).

Classical measurement-driven mechanisms make the control law explicit. In the HSDPA Time Space Priority system, the admissible NRT arrival rate is piecewise linear in total occupancy PRESERVED_PLACEHOLDER_^^^^4all:\4search_query4^^^^:

PRESERVED_PLACEHOLDER_^^^^4all:\4all:\4^^^^

This embeds queue feedback directly into the arrival process (Hanini et al., 2013). AQMRD extends RED by tracking both average queue size and the rate of change of that average, then moves an intermediate threshold PRESERVED_PLACEHOLDER_^^^^4all:\4 OR all:\4^^^^ downward when PRESERVED_PLACEHOLDER_^^^^4all:\4 OR all:\4^^^^ and upward when PRESERVED_PLACEHOLDER_^^^^4all:\44^^^^, thereby making dropping more aggressive when the queue is rising quickly (Karmeshu et al., 2016). FuzzyRTT uses the error PRESERVED_PLACEHOLDER_^^^^4all:\4^^^^5 and PRESERVED_PLACEHOLDER_^^^^4all:\4^^^^6 inside a fuzzy-PI controller, but maintains five RTT-category controllers and interpolates packet-specific drop probability from the RTT carried via Explicit RTT Notification (Boudi et al., 2024).

Delay-based packet AQMs refine not only what is measured but when it is signaled. “Rapid Signalling of Queue Dynamics” argues that ordinary sojourn time is stale at dequeue and proposes the corrected estimate

PRESERVED_PLACEHOLDER_^^^^4all:\4^^^^7

where PRESERVED_PLACEHOLDER_^^^^4all:\4^^^^8 and PRESERVED_PLACEHOLDER_^^^^4all:\4^^^^9 are backlog at enqueue and dequeue, respectively. This is intended to preserve time-domain robustness while reducing signaling lag (&&&^^^^4all:\4^^^^&&&). CAKE takes a different path: it retains an FQ-CoDel core but embeds shaping, overhead compensation, host isolation, DiffServ tiers, NAT awareness, and ACK filtering inside one Linux qdisc, making queue management a coordinated traffic-management stack rather than a single drop function (&&&^^^^4 OR all:\4 OR all:\4^^^^&&&).

Model-driven and learned mechanisms push the control law into an optimizer or a network. PAQMAN defines a discounted SMDP over queue length, arrival-rate state, and RTT-related conditions, then derives an optimal admit/drop policy from a transformed MDP (&&&^^^^4all:\4search_query4^^^^&&&). The call-center scheduler solves an HJB equation for the heavy-traffic diffusion model and uses the learned gradient PRESERVED_PLACEHOLDER_^^^^4 OR all:\4search_query4^^^^ inside a real-time linear program that allocates service activities PRESERVED_PLACEHOLDER_^^^^4 OR all:\4all:\4^^^^ (Ata et al., 10 May 2026). AQM-LLM replaces hand-crafted L4S logic with a distilled action policy over PRESERVED_PLACEHOLDER_^^^^4 OR all:\4 OR all:\4^^^^, while the V^^^^4 OR all:\4^^^^V AoI framework learns a hybrid action PRESERVED_PLACEHOLDER_^^^^4 OR all:\4 OR all:\4^^^^ with PRESERVED_PLACEHOLDER_^^^^4 OR all:\44^^^^ and PRESERVED_PLACEHOLDER_^^^^4 OR all:\4^^^^5, coupling discrete queue management and continuous power control (&&&^^^^4all:\44^^^^&&&, Fang et al., 27 Jan 2026).

A useful synthesis is that richer data do not by themselves determine whether a mechanism is heuristic, optimization-based, or learned. The decisive distinction is where the action map is encoded: in a fixed formula, in a solved control problem, or in trained parameters.

4. System architectures and deployment patterns

A recurring architectural pattern is the separation of intake from service. The medical-chamber system embodies this physically through a patient-side “Patients’ Corner” and a doctor-side “Doctor’s Corner,” with the ATmega^^^^4 OR all:\4 OR all:\4^^^^ handling queue management, sensor acquisition, display control, and synchronization between the two sides (&&&^^^^4search_query4^^^^&&&). Tromino applies the same pattern at cluster scale: users submit tasks to Tromino, dispatchers hold per-framework queues, the manager polls Mesos state, and the scheduler releases work into frameworks according to DRF-aware, demand-aware, or Demand-DRF-aware policies (&&&^^^^4 OR all:\4^^^^&&&). In both cases, queue management becomes a distinct mediation layer rather than a byproduct of downstream service.

Packet-network systems show several deployment strategies. One is integrated queue discipline, exemplified by CAKE, which packages shaping, queueing, and fairness into a single kernel qdisc (&&&^^^^4 OR all:\4 OR all:\4^^^^&&&). Another is wrapper-style augmentation: the ECN-based intelligent AQM keeps CoDel or FQ-CoDel as the underlying discipline but adds an LSTM predictor and Q-learning tuner above it (&&&^^^^4all:\4 OR all:\4^^^^&&&). AQM-LLM adopts a similar wrapper stance in the L4S DualPI^^^^4 OR all:\4^^^^ setting, but replaces rule-based action generation with a state encoder, a backbone LLM, and a one-shot action head (&&&^^^^4all:\44^^^^&&&). FuzzyRTT sits between these extremes: it preserves AQM-style packet dropping but introduces per-category controllers and packet-carried RTT information (Boudi et al., 2024).

Outside packet networking, architectures are often explicitly layered into perception, decision, and actuation. Massimo uses a single-camera image-processing pipeline with YOLO pose estimation, queue-position extraction from hip midpoints, line fitting, confidence-interval outlier detection, force computation, and operator-facing visualization (&&&^^^^4 OR all:\4^^^^&&&). The V^^^^4 OR all:\4^^^^V AoI mechanism uses a bounded two-batch queue model, a hybrid actor with discrete and continuous branches, and a GNN that is computed once per episode from large-scale topology, so decentralized execution can remain lightweight (Fang et al., 27 Jan 2026). In edge computing, the queue-management architecture is tandem: a computation/preprocessing server feeds a transmission server, and the queue discipline changes depending on whether the second stage uses no buffer, one-buffer LCFS with discarding, or preemption (Zou et al., 2019).

The call-center scheduler represents yet another pattern: offline approximation, online optimization. Real data calibrate arrivals, abandonments, service rates, and staffing; a diffusion control problem is solved approximately offline; then a feedforward gradient network and a small linear program determine real-time routing (Ata et al., 10 May 2026). This suggests a broader architectural principle: in high-dimensional service systems, it can be preferable to learn marginal values offline and retain a simple, constrained online optimizer.

5. Empirical evidence across domains

The empirical record is heterogeneous. Some papers report prototype-level feasibility, some provide analytical arguments without experiments, and some present large-scale simulations or calibrated operational studies. Representative quantitative results are summarized below.

Mechanism	Reported result	Scope
Embedded patient system	Temperature accuracy PRESERVED_PLACEHOLDER_^^^^4 OR all:\4^^^^6C over ^^^^4 OR all:\4^^^^5–4^^^^4search_query4^^^^PRESERVED_PLACEHOLDER_ OR all:\4^^^^7C; pulse within PRESERVED_PLACEHOLDER_^^^^4 OR all:\4^^^^8 BPM; button-to-display latency less than ^^^^4all:\4^^^^.^^^^4 OR all:\4^^^^ s	Prototype validation (&&&^^^^4search_query4^^^^&&&)
Intelligent ECN-based AQM	FQ-CoDel occupancy ^^^^4all:\4^^^^.6^^^^4search_query4 average, ^^^^4 OR all:\4^^^^.7^^^^4search_query4^^^^% maximum; CoDel occupancy ^^^^4search_query4^^^^.9^^^^4all:\4 average, ^^^^4 OR all:\4^^^^.^^^^4 OR all:\4search_query4^^^^% maximum	Mininet/Linux emulation (&&&^^^^4all:\4 OR all:\4^^^^&&&)
CAKE	Downstream goodput improves by around ^^^^4all:\4^^^^5%; upload goodput improves by about ^^^^4all:\4search_query4^^^^% with conservative filtering and as much as 4^^^^4search_query4^^^^% with aggressive filtering; latency increases by about 5 ms	Asymmetric ^^^^4 OR all:\4search_query4^^^^/^^^^4all:\4^^^^ Mbps link (&&&^^^^4 OR all:\4 OR all:\4^^^^&&&)
Massimo	YOLOv8: 97.^^^^4 OR all:\4^^^^% accuracy, ^^^^4search_query4^^^^.4^^^^4 OR all:\4^^^^ ms inference time; Linear Regression: 97.47%	Vision-based prototype (&&&^^^^4 OR all:\4^^^^&&&)
AQM-LLM	Llama^^^^4 OR all:\4^^^^-7B training accuracy 97.56%; adapted inference ^^^^4search_query4^^^^.^^^^4search_query4 OR all:\4^^^^9^^^^4all:\4^^^^ s; 64% decrease in GPU memory usage	FreeBSD-^^^^4all:\44^^^^ L4S experiments (&&&^^^^4all:\44^^^^&&&)
Tromino	Waiting-time deviations near cluster average under Demand-DRF aware policy, e.g. PRESERVED_PLACEHOLDER_^^^^4 OR all:\4^^^^9, PRESERVED_PLACEHOLDER_^^^^4 OR all:\4search_query4^^^^, and PRESERVED_PLACEHOLDER_^^^^4 OR all:\4all:\4^^^^ in one workload	8-node Mesos cluster (&&&^^^^4 OR all:\4^^^^&&&)
High-dimensional call-center policy	Optimality gaps ^^^^4search_query4^^^^.7^^^^4all:\4 and ^^^^4search_query4^^^^.75% on ^^^^4 OR all:\4^^^^D tests; 7.^^^^4 OR all:\4 OR all:\4^^^^% better than the best benchmark in a ^^^^4all:\4search_query4search_query4^^^^-dimensional test	Data-calibrated queueing instances (Ata et al., 10 May 2026)

Several additional results sharpen the trade-off picture. The medical-chamber prototype reportedly preserved correct order without data corruption for more than fifty simulated patients, but all entries were volatile and lost on power failure (&&&^^^^4search_query4^^^^&&&). FuzzyRTT reported fairness around 95% under log-normal RTTs, above 8^^^^4search_query4^^^^% even under high congestion in the uniform-RTT case, and near-optimal utilization, while also exhibiting slower stabilization and somewhat higher queueing delay and jitter than some alternatives (Boudi et al., 2024). The call-center policy remained computationally feasible at dimensions up to ^^^^4all:\4search_query4search_query4^^^^ customer classes and consistently matched or exceeded the best benchmark identified in the tested instances (Ata et al., 10 May 2026).

Not all proposals are equally validated. “Rapid Signalling of Queue Dynamics” gives analytical derivations and a signaling-delay table, but explicitly provides no simulation or testbed evidence (&&&^^^^4all:\4^^^^&&&). PAQMAN reports throughput comparable to CoDel with significantly lower queue lengths and delays in simulation, yet its evidence remains simulation-based and its fairness analysis is deferred (&&&^^^^4all:\4search_query4^^^^&&&). The Massimo paper provides promising prototype metrics, but its dataset, camera setup, annotation protocol, and hardware are under-specified, which limits external interpretability (&&&^^^^4 OR all:\4^^^^&&&).

A plausible implication is that empirical maturity correlates strongly with architectural ambition. Lightweight embedded and queue-discipline integrations can often be prototyped directly, whereas high-dimensional or learning-based policies depend more heavily on simulation, emulation, or calibrated surrogates before deployment.

6. Limits, misconceptions, and research directions

The first limitation is semantic. The surveyed works use “data-driven” to describe mechanisms that are record-backed, measurement-informed, model-calibrated, or machine-learned. Treating these as interchangeable obscures important differences in complexity, interpretability, and operational risk. CAKE, for example, is explicitly better described as measurement-informed and heuristic-driven than as learning-based (&&&^^^^4 OR all:\4 OR all:\4^^^^&&&). The embedded patient queue is data-driven because queue entries are structured records, yet the queue discipline is still strict FIFO and the health data do not affect priority (&&&^^^^4search_query4^^^^&&&). Conversely, AQM-LLM and the V^^^^4 OR all:\4^^^^V MAPPO framework are genuinely learned policies whose control law is stored in model parameters rather than in explicit thresholds or recursions (&&&^^^^4all:\44^^^^&&&, Fang et al., 27 Jan 2026).

A second limitation concerns observability and instrumentation. FuzzyRTT depends on Explicit RTT Notification or passive RTT estimation, and its main weakness is slower transient response due to low update frequency (Boudi et al., 2024). Massimo depends on reliable hip keypoints, an interpretable top view, and manageable occlusion, none of which are systematically validated (&&&^^^^4 OR all:\4^^^^&&&). PAQMAN depends on accurate estimation of current arrival-rate state and, in RTT-aware settings, RTT itself (&&&^^^^4all:\4search_query4^^^^&&&). Tromino remains external to Mesos rather than integrated into the master allocator, which simplifies deployment but limits control over the underlying offer mechanism (&&&^^^^4 OR all:\4^^^^&&&).

A third limitation is evidentiary. Some mechanisms remain proof-of-concept or analytically motivated. SAM is explicitly preliminary and under-specifies training data, labels, and SVM configuration (&&&^^^^4all:\4 OR all:\4^^^^&&&). Scaled sojourn time is clearly specified enough to implement, but its claimed benefits remain hypothetical within that paper because no empirical evaluation is provided (&&&^^^^4all:\4^^^^&&&). Even stronger learned systems often rely on carefully controlled settings: AQM-LLM is distilled from trajectories generated by an existing DualPI^^^^4 OR all:\4^^^^ policy, and the V^^^^4 OR all:\4^^^^V framework assumes batch-level all-or-nothing status reconstruction and a two-batch buffer (&&&^^^^4all:\44^^^^&&&, Fang et al., 27 Jan 2026).

The research agenda described in the ML-AQM survey points toward several unresolved problems: computational overhead, sample efficiency, online safety, interpretability, transfer learning, multi-agent learning, Smart Queue Management that integrates shaping and scheduling, real-world benchmarking beyond simple bottlenecks, and deployment in programmable data planes and new network domains (&&&7^^^^4 OR all:\4^^^^&&&). Domain-specific future work in the individual papers aligns with this. The medical kiosk proposes EEPROM or SD-card logging, Bluetooth or Wi-Fi connectivity, mobile waiting-list notifications, and multi-room variants (&&&^^^^4search_query4^^^^&&&). Massimo implies the need for temporal tracking, perspective correction, and real deployment studies (&&&^^^^4 OR all:\4^^^^&&&). The call-center work suggests continued development of scalable approximate control for large sparse skill graphs (Ata et al., 10 May 2026).

The broader trajectory therefore appears to be neither a wholesale replacement of classical queueing rules by black-box learning nor a return to purely handcrafted heuristics. Rather, the literature suggests increasingly explicit queue-state representations, stronger coupling between sensed data and control actions, and hybrid architectures in which measurement, optimization, and learning are combined under operational constraints. In that sense, the data-driven queue management mechanism is less a single algorithm than an evolving design principle: expose queue state in a decision-relevant form, then make control contingent on that state with as much structure as the application and deployment environment can support.