Cost optimization of execution of multi-level deadlineconstrained scientific workflows on clouds

Abstract. This paper introduces a cost optimization model for scien-tific workflows on IaaS clouds such as Amazon EC2 or RackSpace. We assume multiple IaaS clouds with heterogeneous VM instances, with lim-ited number of instances per cloud and hourly billing. Input and output data are stored on a Cloud Object Store such as Amazon S3. Applica-tions are scientific workflows modeled as DAGs as in the Pegasus Work-flow Management System. We assume that tasks in the workflows are grouped into levels of identical tasks. Our model is specified in AMPL modeling language and allows us to minimize the cost of workflow exe-cution under deadline constraints. We present results obtained using our model and the benchmark workflows representing real scientific applica-tions such as Montage, Epigenomics, LIGO. We indicate how this model can be used for scenarios that require resource planning for scientific workflows and their ensembles. Key words: AMPL Optimization; cloud computing; scientific workflows