A Heterogeneous Terascale Computing Cluster for the Development of GPU Optimized High Order Numerical Methods

Abstract : The main goal of the project was to establish a terascale parallel computer cluster on our campus to be shared by the Scientific Computing group, comprised of members of five departments and three colleges at the University of Massachusetts Dartmouth. One of the novel aspects of the proposed computer system is the use of many-core GPUs as hardware accelerators for large scale scientific computation. In The procured system is a 256 CPU IBM iDataPlex system which includes 32 Nvidia Fermi M2050 GPUs as accelerators. The system has been installed, configured and is currently in full operation at the University Data Center. Members of the Scientific Computing group (the investigators and their students) have been successful in 'porting' over their research codes to this new system and are currently in the process of performing detailed tests. Although the system has only been in operation for a few short months, the cluster was already used to perform detailed simulations of the gravitational wave emission from an extreme-mass-ratio black hole binary system. This work resulted in a fast publication in Physical Review. More work is ongoing in the fields of computational mathematics, civil engineering, mechanical engineering, physics, and geophysics.