Performance versus Cost of a Parallel Conjugate Gradient Method in Cloud and Commodity Clusters

Summary Cloud computing is an emerging technology to run HPC applications using computing resources on a pay per use basis. The CG method is a linear solver which is used in many engineering and scientific applications, and is computationally demanding. We implement different approaches of a parallel CG method and compare their performance on different types of platforms: an HPC-optimized cluster, a built heterogeneous cluster, and Amazon cloud. We evaluate the performance of two approaches: broadcast and a ring-based communicationcomputation overlap, and compare to that of National Aeronautics and Space Administration Advanced Supercomputing CG parallel benchmark. We present an evaluation of the performance vis-a-vis cost tradeoff. The results show that, cloud instances suffer from network performance issues, which is revealed by the low performance of the CG method for small problem sizes. An HPC cloud instance type performs the best with relatively less cost than HPC-optimized commodity cluster and other more virtualized cluster instance types, for big problem sizes, while scaling well with increasing problem size. It gives better performance for overlap-based CG method; the performance increases and the cost decreases. Given the emergence of Cloud Computing, the results in this paper analyze performance and cost issues when Clouds are used for CG-based scientific and engineering applications.