Enabling Long-term Fairness in Dynamic Resource Allocation

Abstract

We study the fairness of dynamic resource allocation problem under the α-fairness criterion. We recognize two different fairness objectives that naturally arise in this problem: the well-understood slot-fairness objective that aims to ensure fairness at every timeslot, and the less explored horizon-fairness objective that aims to ensure fairness across utilities accumulated over a time horizon. We argue that horizon-fairness comes at a lower price in terms of social welfare. We study horizon-fairness with the regret as a performance metric and show that vanishing regret cannot be achieved in presence of an unrestricted adversary. We propose restrictions on the adversary's capabilities corresponding to realistic scenarios and an online policy that indeed guarantees vanishing regret under these restrictions. We demonstrate the applicability of the proposed fairness framework to a representative resource management problem considering a virtualized caching system where different caches cooperate to serve content requests.

Fair Resource Allocation in Virtualized O-RAN Platforms

“…Instead, we target a framework that drops these assumptions. The closest to our work is [81], which we extend here in many ways. First, we use a novel optimistic learning fairness algorithm that leverages predictions for the performance and costs.…”

“…performance gains across users over its entire operation; and we include predictions for the unknown (system and user) parameters. Achieving fairness in such dynamic decision models is technically challenging, and previous works are confined to per-slot fairness (which impacts efficiency), with only few exceptions, e.g., [44,55,81]. We overcome this barrier through a saddle-point transformation where in the dual space we track the two fairness metrics.…”

“…Previous works have studied slot fairness criteria [50,82,83] where in each decision round the problem of fairness is tackled independently of the past or future decisions. By definition, the scope of these fairness metrics is limited and it results in higher price of fairness [25,81]. More ambitious approaches attempt to achieve horizon fairness, where the fairness metric is enforced across the entire system operation, not instantly.…”

“…Recent studies on horizon fairness assume the utility functions to be either known or nonadversarial, e.g., study the stochastic version, [6,19,23,30,44,55,81]. Instead, we target a framework that drops these assumptions.…”

“…Unfortunately, off-the-shelf (online) convex optimization algorithms cannot be applied directly on this problem, due to the time-averaging in the argument of functions 𝐹 𝛼 (•) and 𝐹 𝛽 (•), which does not allow the necessary (for these techniques) decomposition over time; see also [1,81]. To tackle this issue, we introduce a proxy function Ψ 𝑡 : Θ × Φ × X ↦ → R with two types of dual variables, 𝜽 ∈ Θ and 𝝓 ∈ Φ, as follows:…”