Online Multi-Commodity Flow with High Demands

This paper deals with the problem of computing, in an online fashion, a maximum benefit multi-commodity flow (\ONMCF), where the flow demands may be bigger than the edge capacities of the network. We present an online, deterministic, centralized, all-or-nothing, bi-criteria algorithm. The competitive ratio of the algorithm is constant, and the algorithm augments the capacities by at most a logarithmic factor. The algorithm can handle two types of flow requests: (i) low demand requests that must be routed along a path, and (ii) high demand requests that may be routed using a multi-path flow. Two extensions are discussed: requests with known durations and machine scheduling.

💡 Research Summary

The paper addresses the online multi‑commodity flow problem (ONMCF) under the realistic assumption that request demands may exceed the capacity of any single edge in the network. Traditional online flow models restrict each demand to be at most the capacity of every traversed edge, which excludes many practical scenarios such as large file transfers, live video streaming, or bulk data replication. To fill this gap, the authors introduce a two‑type request model: (i) low‑demand requests that can be satisfied on a single path, and (ii) high‑demand requests that must be split across multiple paths. Both types are handled in an all‑or‑nothing fashion, meaning a request is either fully accepted (with its entire benefit earned) or completely rejected.

The core contribution is a deterministic, centralized algorithm built on a primal‑dual framework. For each edge a weight w(e) is maintained, and a potential function Φ aggregates the current “price” of using network resources. When a new request r arrives, the algorithm compares the request’s benefit b(r) with the weighted cost of routing it (the sum of w(e) over the chosen path(s) multiplied by Φ). If the benefit outweighs the cost, the request is accepted; the algorithm then updates the weights of all edges used by multiplying them by (1 + ε), where ε is a small constant. This weight‑increase mechanism discourages future over‑use of heavily loaded edges while guaranteeing that the total weight on any edge grows at most logarithmically in the number of edges |E|. The dual side assigns a Lagrange multiplier λ(r) to each request, establishing a lower bound on the optimal offline profit. By carefully coupling the primal updates with the dual variables, the authors prove that the algorithm’s total earned benefit is within a constant factor of the offline optimum—i.e., the competitive ratio is O(1).

A key bi‑criteria guarantee is that the algorithm needs to augment the physical capacities of all edges by at most an O(log |E|) factor to realize the computed fractional flow. This logarithmic augmentation is significantly tighter than the polynomial or linear augmentations required by earlier online flow algorithms, making the solution practically viable for real networks where capacity expansion is costly.

The paper also explores two natural extensions. First, it considers requests with known durations (arrival and departure times). By adding a time dimension, the primal‑dual scheme becomes a two‑dimensional (time‑space) construction, preserving the constant competitive ratio and logarithmic capacity blow‑up. Second, the authors map the problem to classic machine‑scheduling: each request becomes a job, each edge a machine, and the flow amount corresponds to processing time. The same algorithmic ideas apply, yielding a scheduler that respects machine capacities up to a logarithmic factor while achieving a constant‑factor approximation of the optimal weighted throughput.

In the related‑work discussion, the authors contrast their approach with prior online multi‑commodity flow algorithms that rely on randomization, achieve only O(log n) competitive ratios, or require pre‑emptive capacity scaling that can be prohibitive. Their deterministic, constant‑ratio, logarithmic‑augmentation algorithm thus stands out both theoretically and for potential deployment in data‑center networks, content‑distribution systems, and cloud‑based service platforms.

The paper concludes with several avenues for future research: (1) decentralizing the algorithm to eliminate the single‑point‑of‑failure inherent in a centralized controller, (2) extending the model to allow partial acceptance of requests or non‑linear benefit functions, and (3) conducting extensive empirical evaluations on real traffic traces to validate the theoretical guarantees in practice. Overall, the work delivers a robust, theoretically sound, and practically relevant solution to online multi‑commodity flow with high demands, expanding the frontier of online network optimization.