Hybrid Computing Architectures as a Platform for Advanced Multi-scale Computational Methods
Abstract
The Defense University Research Instrumentation Program (DURIP) is designed to improve the capabilities of U.S. Universities to conduct research and to educate scientists and engineers in selected technical areas of importance to national defense. DURIP funding provides for the acquisition of research equipment and instrumentation for this purpose. This proposal is for the purchase of small computer cluster with heterogeneous computing nodes. The P.I., Professor Guglielmo Scovazzi, of Duke University, will use the equipment to augment and enhance research capabilities in the area of Verification and Validation (V&V) of new computational methods, and in basic scalability tests of the resulting codes. Presently, the Mathematical Sciences Division of the Army Research Office (ARO) is sponsoring the PI through a grant titled "Continuous/Discontinuous Variational Multiscale Methods for Variable Density Flows." This project aims at developing algorithms and domain decomposition strategies for subsurface flows, using a multi-scale representation of the solution. By nature, these algorithms have the potential for performance enhancement on heterogeneous computational platforms, which combine conventional computational cores, mathematical co-processors, and GPUs. Although of great potential, this aspect was not pursued in the aforementioned contract by the PI, due to the lack of easy and frequent access to a heterogeneous computing architecture. The PI is aware that some supercomputing clusters in various DOD laboratories and leadership computing facilities already are being equipped with GPUs and mathematical co-processors. However, most of these facilities have restrictions on users who would like to develop software in interactive mode, the preferred approach when intensive software development is sought. The lessons learned from profiling/optimizing the PI s algorithms on the proposed cluster, with various combinations of co-processors/GPUs, are also expected to positively impact the scientific outcomes of the existing project of the PI with ARO, since computational performance, by definition, cannot be abstracted from computing architectures. This proposal hence represents a further extension of the current ARO sponsored PI s research, with the potential to broaden its impact and outcomes.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 12, 2017
- Source ID
- W911NF1510382
Entities
People
- Guglielmo Scovazzi
Organizations
- Army Contracting Command
- Duke University
- United States Army