From Preemptive to Non-preemptive Scheduling Using Rejections

Abstract

We study the classical problem of scheduling a set of independent jobs with release dates on a single machine. There exists a huge literature on the preemptive version of the problem, where the jobs can be interrupted at any moment. However, we focus here on the non-preemptive case, which is harder, but more relevant in practice. For instance, the jobs submitted to actual high performance platforms cannot be interrupted or migrated once they start their execution (due to prohibitive management overhead). We target on the minimization of the total stretch objective, defined as the ratio of the total time a job stays in the system (waiting time plus execution time), normalized by its processing time. Stretch captures the quality of service of a job and the minimum total stretch reflects the fairness between the jobs. So far, there have been only few studies about this problem, especially for the non-preemptive case. Our approach is based to the usage of the classical and efficient for the preemptive case shortest remaining processing time (SRPT) policy as a lower bound. We investigate the (offline) transformation of the SRPT schedule to a non-preemptive schedule subject to a recently introduced resource augmentation model, namely the rejection model according to which we are allowed to reject a small fraction of jobs. Specifically, we propose a $\frac{2}{\epsilon }$-approximation algorithm for the total stretch minimization problem if we allow to reject an $\epsilon$-fraction of the jobs, for any $\epsilon >0$. This result shows that the rejection model is more powerful than the other resource augmentations models studied in the literature, like speed augmentation or machine augmentation, for which non-polynomial or non-scalable results are known. As a byproduct, we present a $\frac{1}{\epsilon }$-approximation algorithm for the total flow-time minimization problem which also rejects at most an $\epsilon$-fraction of jobs.

G. Lucarelli and D. Trystram—This work has been partially supported by the projet Moebus (ANR-13-INFR-0001) funded by ANR.

A. Srivastav—This work has been partially supported by the LabEx PERSYVAL-Lab (ANR-11-LABX-0025-01) funded by the French program “Investissement d’avenir”.