Emergent Mind

Scheduling Servers with Stochastic Bilinear Rewards

(2112.06362)
Published Dec 13, 2021 in cs.LG , cs.DS , and math.OC

Abstract

In this paper, we study scheduling in multi-class, multi-server queueing systems with stochastic rewards of job-server assignments following a bilinear model in feature vectors characterizing jobs and servers. A bilinear model allows capturing pairwise interactions of features of jobs and servers. Our goal is regret minimization for the objective of maximizing cumulative reward of job-server assignments over a time horizon against an oracle policy that has complete information about system parameters, while maintaining queueing system stable and allowing for different job priorities. The scheduling problem we study is motivated by various applications including matching in online platforms, such as crowdsourcing and labour platforms, and cluster computing systems. We study a scheduling algorithm based on weighted proportionally fair allocation criteria augmented with marginal costs for reward maximization, along with a linear bandit algorithm for estimating rewards of job-server assignments. For a baseline setting, in which jobs have identical mean service times, we show that our algorithm has a sub-linear regret, as well as a sub-linear bound on the mean queue length, in the time horizon. We show that similar bounds hold under more general assumptions, allowing for mean service times to be different across job classes and a time-varying set of server classes. We also show stability conditions for distributed iterative algorithms for computing allocations, which is of interest in large-scale system applications. We demonstrate the efficiency of our algorithms by numerical experiments using both synthetic randomly generated data and a real-world cluster computing data trace.

We're not able to analyze this paper right now due to high demand.

Please check back later (sorry!).

Generate a summary of this paper on our Pro plan:

We ran into a problem analyzing this paper.

Newsletter

Get summaries of trending comp sci papers delivered straight to your inbox:

Unsubscribe anytime.