. . "Computer science." . . "Cloud computing." . . "Parallel processing (Electronic computers)" . . . "Addressing scalability and resource provisioning problems for scientific applications on parallel platforms"@en . . . . . . . "Parallel scientific applications often suffer from scalability issues due to design deficiencies and resource provisioning problems when running on emerging parallel platforms such as Cloud. We have designed a scalable approach for LSQR, a widely-used linear system solver, on large-scale high-performance computing platforms and an auto-provisioning tool on Amazon Elastic Compute Cloud. We first present a simple parallel MPI-CUDA approach for the LSQR (called PLSQR). PLSQR's multi-GPU version achieves 3.7x speedup compared with CPU version. However, this parallel approach has the potential, depending on the nonzero structure of the matrix, to have significant communication cost. The communication cost dramatically limits the scalability of the algorithm at large core counts. We then describe a more scalable parallel LSQR algorithm (called SPLSQR) that utilizes the particular nonzero structure of matrices that occur in tomographic problems. In particular, we specially treat the kernel component of the matrix, which is relatively dense with a random structure, and the damping component, which is very sparse and highly structured, separately. The new algorithm has a much more scalable communication volume with a bounded number of communication neighbors regardless of core counts. Experiments on real seismic tomography datasets demonstrates that SPLSQR is scalable on tens of thousands cores and is 38 times faster than implementation in PETSc library. Resource provisioning is one of obstacles when users run parallel scientific applications on Cloud, because it can incur a high monetary expense without careful planning. We have designed a tool called CAP³ (Cloud Auto-Provisioning framework for Parallel Processing) to help a user minimize the expense of running a scientific application on Cloud, while meeting the user-specified job deadline. Given a scientific application, CAP³ automatically profiles the application, builds a model to predict its performance, and infers a proper cluster size that can finish the job within its deadline while minimizing the total cost. To further reduce the cost, CAP³ intelligently chooses the Cloud's reliable on-demand instances or low-cost spot instances, depending on whether or not the remaining time is tight in meeting the application's deadline. Experiments of real parallel scientific applications on Amazon EC2 show that the execution strategy given by CAP³ is cost-effective."@en . "Dissertations, Academic"@en . . . "Computer networks Scalability." . . "University of Wyoming. Computer Science Department." . .