Learning Budget Assignment Policies for Autoscaling Scientific Workflows in the Cloud [JCR]

YISEL GARÍ; DAVID MONGE; CRISTIAN MATEOS; CARLOS GARCÍA GARINO

CLUSTER COMPUTING-THE JOURNAL OF NETWORKS SOFTWARE TOOLS AND APPLICATIONS

SPRINGER

Lugar: Berlin; Año: 2018

1386-7857

Spot instances are extensively used to take advantage of large-scale Cloud infrastructures at lower prices than traditional on-demand instances. Due to the out-of-bid error probability situation, spot instances also represent a trade-off between cost and reliability. Autoscaling scientific workflows in the Cloud considering both spot and on-demand instances presents the challenge of dynamically adjusting the number of instances under each pricing model depending on the workflow needs. Under budget constraints, this adjustment is performed by an assignment policy that determines the suitable proportion of the available budget intended for each model. Given uncertainty sources derived from inherent performance variability of Clouds and the unpredictability of out-of-bid errors, we formalize the described problem as a Markov Decision Process and derive adaptive policies by learning from the experiences of other baseline policies. Experiments over four well-known scientific workflows, demonstrate that learned policies outperform all the baseline policies considering the aggregated relative percentage difference of makespan and execution cost. These promising results encourage the future study of new strategies aiming to find optimal budget policies applied to the execution of workflows in the Cloud./* Layout-provided Styles */div.abstract {margin-top: 0.7ex;margin-bottom: 0.7ex;text-align: left;}span.abstract_label {font-weight: bold;}