Practical Divisible Load Scheduling on Grid Platforms with APST-DV

Divisible load applications consist of a load, that is input data and associated computation, that can be divided arbitrarily into independent pieces. Such applications arise in many fields and are ideally suited to a master-worker execution, but they pose several scheduling challenges. While the "Divisible Load Scheduling" (DLS) problem has been studied extensively from a theoretical standpoint, in this paper we focus on practical issues: we extend a production Grid application execution environment, APST, to support divisible load applications; we implement previously proposed DLS algorithms as part of APST; we evaluate and compare these algorithms on a real-world two-cluster platform; we show in a case study how a user can easily and effectively run a real-world divisible load application; and we uncover several issues that are critical for using DLS theory in practice. To the best of our knowledge the software resulting from this work, APST-DV, is the first usable and generic tool for deploying divisible load applications on distributed computing platforms.