Running Scientific Codes on Amazon EC2: a Performance Analysis of Five High-end Instances

Resumo

Amazon Web Services (AWS) is a well-known public Infrastructure-as-a-Service (IaaS) provider whose Elastic Computing Cloud (EC2) offering includes some instances, known as cluster instances, aimed at High-Performance Computing (HPC) applications. In previous work, authors have shown that the scalability of HPC communication-intensive applications does not benefit from using higher computational power cluster instances as much as it could be expected. Cost analysis recommends using lower computational power cluster instances unless high memory requirements preclude their use. Moreover, it has been observed that scalability is very poor when more than one instance is used due to network virtualization overhead. Based on those results, this paper gives more insight into the performance of running scientific applications on the Amazon EC2 platform evaluating five (of which two have been recently released) of the higher computational power instances in terms of single instance performance, intra-VM (Virtual Machine) scalability and cost-efficiency. The evaluation has been carried out using both an HPC benchmark suite and a real High-Troughput Computing (HTC) application.