Scheduling with Outliers

TL;DR

This paper studies scheduling problems with outliers, where only a subset of jobs meeting a profit target are scheduled to optimize various objectives, and provides LP-based approximation algorithms for these problems.

Contribution

It introduces LP-based approximation algorithms for scheduling with outliers across multiple objectives, including new bounds and integrality gap results.

Findings

Logarithmic approximation for average flow time with unit profits

Constant factor approximation for weighted completion time on unrelated machines

3-approximation for generalized assignment with outliers

Abstract

In classical scheduling problems, we are given jobs and machines, and have to schedule all the jobs to minimize some objective function. What if each job has a specified profit, and we are no longer required to process all jobs -- we can schedule any subset of jobs whose total profit is at least a (hard) target profit requirement, while still approximately minimizing the objective function? We refer to this class of problems as scheduling with outliers. This model was initiated by Charikar and Khuller (SODA'06) on the minimum max-response time in broadcast scheduling. We consider three other well-studied scheduling objectives: the generalized assignment problem, average weighted completion time, and average flow time, and provide LP-based approximation algorithms for them. For the minimum average flow time problem on identical machines, we give a logarithmic approximation algorithm for the case of unit profits based on rounding an LP relaxation; we also show a matching integrality gap. For the average weighted completion time problem on unrelated machines, we give a constant factor approximation. The algorithm is based on randomized rounding of the time-indexed LP relaxation strengthened by the knapsack-cover inequalities. For the generalized assignment problem with outliers, we give a simple reduction to GAP without outliers to obtain an algorithm whose makespan is within 3 times the optimum makespan, and whose cost is at most (1 + \epsilon) times the optimal cost.