Throughput in Asynchronous Networks

We introduce a new, “worst-case” model for an asynchronous communication network and investigate the simplest (yet central) task in this model, namely the feasibility of end-to-end routing. Motivated by the question of how successful a protocol can hope to perform in a network whose reliability is guaranteed by as few assumptions as possible, we combine the main “unreliability” features encountered in network models in the literature, allowing our model to exhibit all of these characteristics simultaneously. In particular, our model captures networks that exhibit the following properties: 1) On-line; 2) Dynamic Topology; 3)Distributed/Local Control 4) Asynchronous Communication; 5) (Polynomially) Bounded Memory; 6) No Minimal Connectivity Assumptions. In the confines of this network, we evaluate throughput performance and prove matching upper and lower bounds. In particular, using competitive analysis (perhaps somewhat surprisingly) we prove that the optimal competitive ratio of any on-line protocol is 1/n (where n is the number of nodes in the network), and then we describe a specific protocol and prove that it is n-competitive. The model we describe in the paper and for which we achieve the above matching upper and lower bounds for throughput represents the “worst-case” network, in that it makes no reliability assumptions. In many practical applications, the optimal competitive ratio of 1/n may be unacceptable, and consequently stronger assumptions must be imposed on the network to improve performance. However, we believe that a fundamental starting point to understanding which assumptions are necessary to impose on a network model, given some desired throughput performance, is to understand what is achievable in the worst case for the simplest task (namely end-to-end routing).

💡 Research Summary

The paper introduces a rigorously defined “worst‑case” model for asynchronous communication networks that makes virtually no reliability assumptions. The model simultaneously incorporates six key sources of unreliability that appear across the literature: (1) online operation (inputs are revealed only at run time), (2) dynamic topology (nodes and links may appear or disappear arbitrarily), (3) distributed/local control (each node knows only its immediate neighbors), (4) asynchronous message delivery (no global clock, unbounded delays), (5) polynomially bounded memory at each node, and (6) no minimal connectivity guarantee (the network may become completely disconnected). Within this hostile environment the authors study the most fundamental networking task—end‑to‑end packet routing—and evaluate its throughput using competitive analysis.

Throughput is defined as the number of packets successfully delivered from source to destination per unit time. The competitive ratio r of an online protocol P is the smallest value such that for every possible input sequence, throughput(P) ≥ (1/r)·throughput(OPT), where OPT is an optimal offline algorithm that knows the entire future. The authors first construct an adversarial scheduler that forces any protocol into a “star” topology: a central hub connected to n − 1 leaf nodes, each generating an independent stream of packets. The adversary can interleave packet injections and edge deletions so that each leaf’s flow must pass through the hub, and the hub’s limited memory forces it to allocate at most a 1/n fraction of its capacity to each leaf. This construction yields a lower bound: no online protocol can achieve a competitive ratio better than 1/n.

To match this lower bound, the paper proposes a concrete protocol called Uniform Distribution Routing (UDR). UDR operates with only local information: each node maintains a FIFO queue of bounded size (O(poly(n))) and, upon each transmission opportunity, forwards the head‑of‑queue packet to its neighbors in a round‑robin fashion. By evenly spreading traffic across all incident edges, UDR guarantees that, even under the worst‑case adversarial schedule, each node contributes at least a 1/n share of the total possible flow. The authors prove, using a combination of potential‑function arguments and Markov‑chain analysis, that UDR is n‑competitive; that is, its throughput is within a factor n of the optimal offline benchmark. Consequently, the upper bound matches the previously established lower bound, establishing that the optimal competitive ratio for this model is exactly 1/n.

The significance of these results is twofold. First, they provide a precise characterization of what is achievable when no reliability assumptions are made—a baseline that was missing from prior work, which typically assumed at least one stabilizing property (e.g., fixed connectivity, bounded delay, or global routing tables). Second, the extremely low competitive ratio (1/n) demonstrates that, in practice, additional assumptions are indispensable if a system must deliver reasonable throughput. The paper therefore serves as a diagnostic tool: by identifying which assumptions (such as a minimal connectivity guarantee, bounded latency, or synchronized rounds) are needed to improve the competitive ratio, designers can make informed trade‑offs between robustness and performance.

The authors conclude with several avenues for future research. They suggest a hierarchical analysis where incremental assumptions are added to the model and the resulting competitive ratios are quantified. They also propose extending the framework to multi‑objective settings that incorporate latency, energy consumption, or security constraints, and to evaluate how these objectives interact with throughput under adversarial dynamics. Finally, they recommend empirical validation of UDR through simulation and real‑world testbeds, as well as exploring stronger adversarial models (e.g., Sybil or Eclipse attacks) to understand the security‑throughput trade‑off in the worst‑case regime.

In summary, the paper establishes that in a fully asynchronous, dynamic, and memory‑constrained network with no connectivity guarantees, the best possible competitive throughput is 1/n, and it provides a simple, locally implementable protocol that attains this bound. This work lays a theoretical foundation for systematically assessing which additional network assumptions are necessary to achieve higher performance in realistic distributed systems.