A Predictive and Evolutionary Approach for Cost-Effective and Deadline-Constrained Workflow Scheduling Over Distributed IaaS Clouds

Abstract

Clouds provide highly elastic resource provisioning styles through which scientific workflows are allowed to acquire desired resources ahead of the execution and build required software environment on virtual machines (VMs). However, various challenges for cloud workflow, especially its optimal scheduling, are yet to be addressed. Traditional approaches mainly consider VMs to be with non-fluctuating, time-invariant, stochastic, or bounded performance. This work describes workflows to be deployed and executed over distributed infrastructure-as-a-service clouds with time-varying performance of VMs and is aimed at reducing the execution cost of workflow while meeting deadline constraints. For this purpose, the authors employ time-series-based prediction approaches to capture dynamic performance fluctuations, feed an evolutionary algorithm with predicted performance information, and generate schedules at real-time. A case study based on multiple randomly-generated workflow templates and third-party commercial clouds shows that their proposed approach outperforms traditional ones.