Authors:  He Huang, Liqiang Wang, University of Wyoming
                 Byungchul Tak, Long Wang, Chunqiang Tang, IBM T.J. Watson Research Center

Traditionally, scientific applications require massive computing resources, hence usually run on dedicated high performance computing (HPC) clusters. The recent advances of Cloud Computing have made utility computing widely available. Infrastructure as a service (IaaS), such as Amazon EC2, makes it possible to run parallel scientific applications in a way of pay-as- you-go. Determining a proper size of virtual cluster is critical for parallel scientific applications in cloud. A customer is usually not clear what a proper size is to meet the requirement. A smaller size may result in very long execution time, which may exceed the customer’s time constrain. A larger size could reduce execution time but may be unnecessary because parallel applications usually do not scale after a sufficient large size. In addition, larger size can potentially cause more costs. We propose an automatic framework of resource provisioning and sizing for HPC applications in cloud environment. We use a light weight performance model to obtain scalability features of parallel scientific applications. Based on these scalability features, our evaluation model can determine a proper size of cloud virtual cluster according to customer requirement.