Early prediction of the cost of cloud usage for HPC applications

Abstract

After a decade of diffusion, cloud computing has received wide acceptance, but it is not yet attractive for the HPC community. Clouds could be a cost-effective alternative to clusters and supercomputers, providing economy of scale, elasticity, flexibility, and easy customization. Unfortunately, most clouds are optimized for running business applications, not for HPC. However, they can be profitably used to run small-scale parallelism codes. This paper presents a framework built on the top of a cloud-aware programming platform (mOSAIC) for the development of bag-of-tasks scientific applications. The framework integrates a cloud-based simulation environment able to predict the behavior of the developed applications. Simulations enable the developer to predict at an early development stage performance and cloud resource usage, and so the infrastructure lease cost on a public cloud. The paper sketches the framework organization and presents the approach followed for the performance simulation of applica-tions, focusing on a software development methodology that hinges on early performance prediction. After showing the results of some validation tests of simulation accuracy, an example of early performance prediction is presented.