Traffic dynamics in scale-free networks with limited packet-delivering capacity

We propose a limited packet-delivering capacity model for traffic dynamics in scale-free networks. In this model, the total node’s packet-delivering capacity is fixed, and the allocation of packet-delivering capacity on node $i$ is proportional to $k_{i}^{\phi}$, where $k_{i}$ is the degree of node $i$ and $\phi$ is a adjustable parameter. We have applied this model on the shortest path routing strategy as well as the local routing strategy, and found that there exists an optimal value of parameter $\phi$ leading to the maximal network capacity under both routing strategies. We provide some explanations for the emergence of optimal $\phi$.

💡 Research Summary

The paper addresses the problem of traffic congestion in large communication networks such as the Internet and the World Wide Web, focusing on the realistic constraint that the total packet‑delivering capacity of all nodes is limited. Using the Barabási‑Albert (BA) model to generate scale‑free networks, the authors introduce a traffic model in which, at each discrete time step, R packets are generated with random source‑destination pairs. Each node i can forward at most C_i packets per step, where the set {C_i} is constrained by a fixed average capacity ⟨C⟩. The key novelty is the allocation rule

C_i = N⟨C⟩ · k_i^φ / ∑_j k_j^φ,

where k_i is the degree of node i and φ is a tunable parameter. φ>0 gives larger capacity to high‑degree nodes, φ<0 does the opposite, and φ=0 corresponds to uniform capacity.

Two routing strategies are examined. (1) Shortest‑path routing, where each packet follows a geodesic between source and destination. (2) Local routing, where a node first checks its immediate neighbours for the destination; if not found, it forwards the packet to neighbour i with probability Π_i = k_i^α / ∑_j k_j^α, where α controls the bias toward high‑degree neighbours.

The system’s performance is quantified by the order parameter η(R) = lim_{t→∞} (1/R)⟨ΔN_p/Δt⟩, where N_p(t) is the total number of packets present at time t. η=0 indicates a free‑flow phase, while η>0 signals a congested phase. The critical packet generation rate R_c, at which η first becomes positive, serves as the measure of network capacity.

Simulation results on BA networks (N=1000, m_0=m=3 or 4) reveal that R_c is not monotonic in φ. For both routing strategies there exists an optimal φ_opt that maximizes R_c. In the shortest‑path case, the authors relate this to betweenness centrality g(k), which scales as g(k)∝k^μ with μ≈1.33 for the considered networks. When φ≈μ, the capacity distribution C_i∝k_i^φ matches the load distribution imposed by shortest‑path traffic, leading to balanced utilization and the highest R_c. If φ is too small, high‑degree nodes become bottlenecks; if φ is too large, many low‑degree nodes are under‑provisioned, both causing early congestion.

For the local routing scheme, prior work shows that in the free‑flow regime the average queue length obeys n(k)∝k^{1+α}. Consequently, setting φ≈1+α aligns capacity with expected load. Simulations confirm φ_opt≈1+α across a range of α values, and the maximal R_c exhibits a non‑monotonic dependence on α, peaking near α≈0.2. Detailed inspection of queue lengths in the congested regime demonstrates that φ below the optimum leads to excessive queues on hubs, while φ above the optimum causes queues to accumulate on peripheral nodes.

The authors conclude that, under a fixed total delivering capacity, a simple power‑law allocation of node capacities controlled by a single exponent φ can dramatically improve traffic handling. The optimal exponent is dictated by the routing‑induced load distribution: φ_opt≈μ for shortest‑path routing and φ_opt≈1+α for the local biased random walk. This insight provides a practical guideline for designing cost‑effective communication infrastructures where hardware resources (bandwidth, processing power) must be distributed among heterogeneous nodes.