Scheduling in the high uncertainty heavy traffic regime

TL;DR

This paper introduces a model uncertainty approach to heavy traffic queueing systems, analyzing a stochastic game between a scheduler and an adversary under high uncertainty, and derives optimal policies and limits.

Contribution

It develops a novel heavy traffic asymptotic framework accommodating large distributional disturbances using only first two moments, and fully solves the associated stochastic game.

Findings

Optimal scheduling policy is an index rule.

Adversary's optimal strategy depends on current workload.

Limit dynamics are characterized by a discontinuous SDE.

Abstract

We propose a model uncertainty approach to heavy traffic asymptotics that allows for a high level of uncertainty. That is, the uncertainty classes of underlying distributions accommodate disturbances that are of order 1 at the usual diffusion scale, as opposed to asymptotically vanishing disturbances studied previously in relation to heavy traffic. A main advantage of the approach is that the invariance principle underlying diffusion limits makes it possible to define uncertainty classes in terms of the first two moments only. The model we consider is a single server queue with multiple job types. The problem is formulated as a zero sum stochastic game played between the system controller, who determines scheduling and attempts to minimize an expected linear holding cost, and an adversary, who dynamically controls the service time distributions of arriving jobs, and attempts to maximize the cost. The heavy traffic asymptotics of the game are fully solved. It is shown that an asymptotically optimal policy for the system controller is to prioritize according to an index rule and for the adversary it is to select distributions based on the system's current workload. The workload-to-distribution feedback mapping is determined by an HJB equation, which also characterizes the game's limit value. Unlike in the vast majority of results in the heavy traffic theory, and as a direct consequence of the diffusive size disturbances, the limiting dynamics under asymptotically optimal play are captured by a stochastic differential equation where both the drift and the diffusion coefficients may be discontinuous.