Threshold-p Routing Strategies

• Threshold-p routing is defined as selecting network paths or system actions when performance metrics like cost, delay, and throughput meet predefined thresholds.
• It employs mathematical models and multi-criteria decision-making frameworks, such as ISOCOV, to rigorously assess and rank feasible routing options.
• Applications span dynamic relay routing, congestion control, and adaptive load balancing in networking, queueing systems, and machine learning architectures.

Threshold-p Routing is a class of routing policies and decision mechanisms that select paths, nodes, or system behaviors based on whether key metrics cross certain thresholds. Across networking, queueing theory, optimization, and machine learning, threshold-p frameworks rigorously encode performance guarantees, trade-offs, and system constraints through mathematically-defined threshold vectors, rules, or distributions. This article provides a comprehensive analysis of threshold-p routing, covering both its multi-metric and single-metric instantiations, operational principles, algorithmic structures, representative case studies, and theoretical properties.

1. Mathematical Formulation and Model Structure

Threshold-p routing in its most general form involves selecting paths in a network that meet predefined thresholds on a collection of $p$ performance metrics. Formally, consider a network $G=(V,E)$, and let $P$ denote a simple path from source $s\in V$ to destination $t\in V$. For each such path, $p$ real-valued metrics $m_1,\ldots,m_p$ are defined, such as cost, delay, loss, throughput, and jitter. Each metric $m_j$ is aggregated over the path using an appropriate operator (e.g., summation for additive costs, minimum for bottleneck capacities).

A threshold vector $T=(\tau_1,\ldots,\tau_p)$ encodes the required upper or lower bounds for these metrics. The feasible set of paths consists of those satisfying the threshold-p constraints: $m_j(P) \leq \tau_j \quad \forall j=1,\ldots,p$ This single definition underpins a wide spectrum of threshold-p routing systems, from constrained shortest-path problems to modern multi-criteria decision making in SDN and IoT networks (Bouakouk et al., 2022).

The threshold-p routing paradigm may also appear in simpler, single-threshold policies, such as queue-length-based switching in dynamic systems, or probability-mass-based selection rules in machine learning (e.g., Top-p coverage in Mixture-of-Experts).

2. Multi-Criteria Threshold-p Routing with MCDM

When routing must simultaneously satisfy multiple performance metrics and optimize trade-offs among them, threshold-p routing couples feasibility (threshold admissibility) with Multi-Criteria Decision Making (MCDM). In (Bouakouk et al., 2022), the ISOCOV-based MCDM framework operationalizes this as follows:

• Candidate paths are enumerated and evaluated on each metric, forming a decision matrix $D\in \mathbb{R}^{m\times p}$.
• For each metric $j$, a threshold interval $C_j=[A_j,B_j]$ is imposed, with a "hard" or "soft" flag indicating strict versus flexible satisfaction.
• Satisfaction degrees $f_{ij}\in(0,1]$ quantify how well path $i$ matches the threshold on metric $j$.
• Normalization, weighting, and aggregation produce adjusted scores for each path.
• The ISOCOV closeness coefficient $C_i$ combines distance to "Positive Ideal Solution" and "Negative Ideal Solution," favoring paths that strictly or approximately satisfy all metric thresholds.

This process produces a ranking of all threshold-feasible paths, with hard-threshold satisfaction strictly enforced ($C_i<0$ for any violation) or, with soft constraints, admitting controlled trade-offs. The overall computational cost is $O(m\cdot p + m\log m)$ for $m$ candidate paths and $p$ metrics. The dominant cost in practical settings is path enumeration.

3. Classical and Control-Theoretic Threshold Policies

A broad class of threshold-p routing arises in stochastic control and queueing systems, where routing (or switching) is governed by whether a system state exceeds a critical threshold. Representative examples include:

• Dynamic relay routing in wireless networks: The optimal policy for relaying or direct transmission is defined by a queue-length threshold $p$, characterized analytically by the solution to a monotonicity condition in the Bellman equations. For instance, with arrival rate $\lambda$ and service rates $\mu_d,\mu_r$, the optimal threshold is

$$p = \left\lfloor \frac{\log\left[\frac{\mu_r-\lambda}{\mu_d-\lambda}\right]}{\log(\mu_r/\mu_d)} \right\rfloor$$

Routing switches from relay to direct when the buffer size exceeds $p$ (Cohen et al., 2016).

• Decentralized threshold routing in queuing networks: With symmetric, coupled queue systems, decentralized controllers exchange implicit information by routing one customer if their post-arrival queue exceeds the shared threshold (midpoint of the estimate support). This policy provably minimizes convex holding costs in both finite and infinite horizons under matching initial conditions (Ouyang et al., 2014).
• Threshold-based route discovery in mobile ad hoc networks: The trade-off between frequent, costly route discovery and throughput loss due to route staleness is optimally managed using a single throughput threshold $T$. Discovery is triggered only when realized throughput $r_t<T$, with $T$ determined by a renewal-reward approximation:

$T \approx (1-\phi^2)E[R^*]$

where $\phi$ is the fraction of slot lost to discovery, and $E[R^*]$ is expected best-case rate (Patra et al., 2010).

4. Threshold-p Routing in Modern Machine Learning Systems

Threshold-p ideas appear in neural network architectures, particularly sparse mixture-of-experts (MoE) models:

• Top-p (nucleus) routing in MoE: Each token activates a minimal set of experts such that the cumulative routing probability exceeds $p$. Traditional fixed-p methods lack strict control over average activated experts, causing unpredictable compute and sensitivity.
• Dynamic Top-p (DTop-p) routing: (Jin et al., 16 Dec 2025) introduces a Proportional-Integral controller to adaptively adjust the Top-p threshold, precisely aligning average expert activations to a user-specified budget $T$. Per-layer learnable temperature parameters further allow layer-specific sparsity patterns, dramatically outperforming Top-k and fixed-p strategies in both NLP and vision foundation model training.

The DTop-p mechanism enforces:

• Strong compute-activation invariance (via PI feedback).
• Layer-adaptivity (via normalization).
• Robust convergence and downstream performance (multi-benchmark superiority).

5. Threshold-based Routing for Congestion and Admission Control

Threshold-p routing frameworks are central in admission and congestion control systems. For example:

• ACL-based threshold rerouting: Routers enforce per-link congestion thresholds in extended ACL rules (Mojaveri et al., 2021). When measured link load $L_\ell\geq\tau$, forwarding decisions switch to alternate routes or drop low-priority packets. The choice of threshold $\tau$ directly trades between packet loss and reroute frequency. Empirical results confirm substantial reduction in drop percentages as the threshold is tuned below $100\%$ utilization.
• Service systems with priority/routing thresholds: In parallel server systems offering both fixed-price FIFO and auction-based highest-bidder-first (HBF) queues, equilibrium routing is characterized by two thresholds $\theta_1,\theta_2$ that segment customer valuations. Customers with $V<\theta_1$ or $V>\theta_2$ choose HBF; those in $(\theta_1,\theta_2)$ choose FIFO. Thresholds are pinned by Wardrop equilibrium conditions, with system parameters (e.g., price, demand) tuning the size of threshold-determined segments (Bodas et al., 2016).

6. Algorithmic Implementation and Case Studies

Threshold-p routing policies take multiple algorithmic forms, but share common procedural elements:

• System state observation and metric computation (either instantaneous, e.g., queue length, or aggregate, e.g., average path loss).
• Comparison to precomputed or adaptively-updated threshold(s).
• Decision: selection of path, action, or routing table based on the threshold rule.
• For multi-criteria settings, an explicit ranking (e.g., via ISOCOV) of all threshold-feasible candidates.

Representative case studies illustrate the practical performance of threshold-p routing:

• On a 7-node, 27-path network, ISOCOV MCDM with hard threshold-p constraints selects robust paths meeting all QoS thresholds, outperforming classical approaches that lack interval-based constraints. With soft thresholding, the ranking mechanism allows performance-outlier paths to rise if they marginally violate less critical criteria (Bouakouk et al., 2022).
• In fat-tree datacenter networks, threshold-based two-choice randomized routing rules minimize both uplink congestion and downlink collision, attaining doubly-exponential queue-size tail bounds and superior queue and latency profiles versus baseline randomized routing (Wang et al., 2017). The per-switch threshold $T$ moderates traffic redistribution to avoid excessive route flapping.

7. Parameter Tuning, Trade-offs, and Theoretical Insights

The performance and optimality of threshold-p routing critically depend on the choice and handling of thresholds:

• Hard vs. soft enforcement: Hard thresholds enforce strict admittance and may exclude high-performing options; soft thresholds facilitate utility trade-offs at the risk of relaxing guarantees.
• Parameter estimation and adaptation: Where path or system statistics drift, online estimation and dynamic threshold adjustment (e.g., PI controllers (Jin et al., 16 Dec 2025), re-estimation from sliding windows (Cohen et al., 2016)) are essential for sustained optimality.
• Trade-off surfaces: Lowering thresholds reduces congestion and drops but may trigger more frequent rerouting (increasing jitter or system overhead). Raising thresholds increases single-mode operation but risks QoS degradation under dynamic load (Mojaveri et al., 2021).
• Scalability and complexity: For practical threshold-p routing, enumeration is often constrained to $k$-shortest or otherwise limited sets, keeping evaluation tractable [$O(m\cdot p + m\log m)$ in the multi-metric case].

Threshold-p routing thus provides a mathematically principled, operationally robust basis for configuring resource allocation, routing, and optimization systems under multi-dimensional, sometimes competing, metrics and evolving system conditions.