Network Congestion Control with Markovian Multipath Routing

In this paper we consider an integrated model for TCP/IP protocols with multipath routing. The model combines a Network Utility Maximization for rate control based on end-to-end queuing delays, with a Markovian Traffic Equilibrium for routing based on total expected delays. We prove the existence of a unique equilibrium state which is characterized as the solution of an unconstrained strictly convex program. A distributed algorithm for solving this optimization problem is proposed, with a brief discussion of how it can be implemented by adapting the current Internet protocols.

💡 Research Summary

The paper addresses the long‑standing separation between routing (IP layer) and congestion control (TCP layer) in the Internet architecture. While IP typically provides a static single‑path routing that updates slowly, TCP adjusts the transmission window rapidly based on congestion signals such as queueing delay or packet loss. This separation can cause instability when routing metrics depend on congestion, leading to route flaps and sub‑optimal utilization of available bandwidth. To overcome these issues, the authors propose a unified analytical framework that simultaneously models rate control and multipath routing as a single convex optimization problem.

The rate‑control component is built on the well‑known Network Utility Maximization (NUM) formulation. Each source k adjusts its sending rate x_k as a decreasing function f_k of the aggregate queueing delay q_k experienced along its path. The equilibrium condition x_k = f_k(q_k) can be expressed as the KKT conditions of a strictly convex program that maximizes the sum of source utilities subject to link capacity constraints. The utility functions are primitives of the inverse of f_k, while the link cost functions are primitives of the queueing‑delay functions ρ_a(w_a).

The routing component is first described by Wardrop’s equilibrium, where only minimum‑delay paths carry flow. To capture stochastic variations in link delays, the authors introduce a Markovian Traffic Equilibrium (MTE). They assume that each link’s random delay is the sum of a deterministic expected delay λ_a and an i.i.d. Gumbel noise. This leads to a Logit choice model for path splitting, but because the number of paths can be exponential, a direct implementation is infeasible. Instead, they formulate routing as a stochastic dynamic program: at each node i a packet chooses the outgoing link a that minimizes the random quantity ˜z_{ka}=˜λ_a+τ_{kj_a}, where τ_{kj} is the expected downstream delay to the destination. The resulting transition probabilities define a Markov chain whose invariant measure yields the flow distribution across links.

By coupling the NUM and MTE models, the authors derive the Markovian Network Utility Maximization (MNUM) problem. The decision variables collapse to the link congestion prices λ = (λ_a)_{a∈A}. The objective function

Φ(λ) = Σ_a ∫_0^{λ_a} s_a^{-1}(y) dy – Σ_k F_k(τ_k(λ))

is strictly convex and unconstrained, where s_a^{-1} is the inverse of the total delay function, τ_k(λ) is the minimal expected end‑to‑end delay for source k given the current λ, and F_k is the primitive of the source’s rate‑control function. The authors prove that Φ has a unique minimizer λ*; the corresponding λ* together with the induced rates x_k = f_k(τ_k(λ*)) and flow splits satisfy all original equilibrium equations. This variational characterization provides both existence and uniqueness of the joint routing‑rate equilibrium.

A distributed algorithm is proposed to solve the MNUM problem. Each link measures its current load w_a, computes the gradient of its local term ∂Φ/∂λ_a = s_a^{-1}(λ_a) – Σ_{k: a∈path_k} ∂F_k/∂τ_k·∂τ_k/∂λ_a, and updates λ_a via a gradient‑ascent (or dual‑decomposition) step. Sources receive the updated λ values (e.g., via existing RTT feedback mechanisms), recompute their target delay τ_k(λ), adjust their window size according to x_k = f_k(τ_k), and update their splitting probabilities according to the Markovian routing rule. The algorithm requires only local measurements and occasional price broadcasts, making it compatible with current TCP‑Vegas style delay‑based congestion control and with incremental changes to IP routing tables.

The paper discusses practical considerations, such as the need for accurate estimation of expected delays, the assumption of Gumbel‑distributed noise, and the impact of network topology (e.g., Braess’s paradox). It acknowledges that while the convex formulation guarantees optimality under the model assumptions, real networks may exhibit correlated delays, non‑stationary traffic, and measurement noise that violate these assumptions. Consequently, the authors suggest future work on robust extensions, experimental validation, and integration with emerging protocols like Multipath TCP (MPTCP) and Software‑Defined Networking (SDN) controllers.

In summary, the work makes a significant theoretical contribution by unifying rate control and multipath routing within a single strictly convex optimization framework, proving the existence of a unique equilibrium, and outlining a feasible distributed algorithm that could be adapted to existing Internet protocols. It opens a promising avenue for cross‑layer design that leverages both congestion‑aware routing and delay‑based congestion control to improve throughput, fairness, and stability in modern packet‑switched networks.