Adaptive Fair Scheduler: Fairness in Presence of Disturbances

Abstract

The problem of allocating resources over time to different demands in a "fair" way is present in many application domains. If the resource can be allocated with an arbitrarily fine granularity at no cost, then any type of resource allocation can be achieved (this scheme is called fluid for its resemblance to water). Instead, if the resource has some coarse granularity, then the fluid resource allocation can only be approximated. The notion of lag measures the deviation between the fluid schedule and the real schedule which can be actually achieved. In this paper, we propose the Adaptive Fair Scheduler (AFS), which allocates resources over time and guarantees a target service rate to a set of applications. AFS is capable of achieving a bounded lag in presence of time overhead at scheduling decision instants, and uncertainties in the resource allocation. Thanks to its generality, AFS can be applied to many different application domains. Reconfigurable computing, scheduling of heterogeneous units, and multiprocessor scheduling are some notable examples.