Acceleration of Commercial Low-Rank Matrix Solver via SLATE
Performing accurate simulations of large- and multi-scale electromagnetics problems has far-reaching implications. The same physics governs applications as diverse as rare earth content reduction in DC motor design; wireless propagation in dense urban environments and novel antenna design, both key issues in the pending 5G communications boom; electromagnetic sounding for subsurface exo-terrestrial interrogation; burgeoning RF-based medical imaging modalities; and exploration of intra- and inter-cellular quasistatic and RF signaling. With such diversity of simulation problems, a scalable, high performance computational electromagnetics (CEM) software package will provide state-of-the-art design and analysis capability to commercial, medical, and government users. IERUS Technologies has developed a commercial CEM software package called V-Lox for performing electrically large-scale simulations using novel matrix compression techniques with multi-GPU hardware acceleration. V-Lox exploits the approximate low-rank block structure of the underlying surface integral equation formulation to achieve signi cant reduction in the memory and computational demands for electrically large and complex problems with millions of degrees of freedom (DoF). However, the commercial version of V-Lox is currently limited to a single workstation/server and can simulate problems with total electrical size of ~ 100 - 200λ (wavelengths). To achieve next-generation computational capability on peta- and exascale platforms, an HPC-targeted solution is required. With our partner Innovative Computing Lab (ICL) at the University of Tennessee-Knoxville, we have previously shown the potential of an HPC-enabled software package leveraging the latest in massively parallel, multi-GPU solution capability. During this program, IERUS and ICL have combined the unique CEM solution capabilities of V-Lox with the US Department of Energy (DOE)-funded project Software for Linear Algebra Targeting Exascale (SLATE). Our work combined the existing capabilities of the two software packages and, importantly, extended the application of SLATE to target modern low-rank-exploiting linear algebra algorithms. V-Lox produces matrices with non-uniform compressed and uncompressed block structure on a large scale. Utilizing the matrix storage abstraction capability inherent to SLATE, the team extended SLATE to handle this class of matrices while maintaining high performance multi-CPU resources. By incorporating this modifi ed version of SLATE into V-Lox, the team has ensured that all developments were pursued with industrial-scale robustness and tested on highly realistic problems. The team successful tested the performance of the modi ed software to address problems relevant to ion trap design for quantum computing. Testing addressed performance evaluation on modest and high-performance computational resources using up to 64 multi-core processors. Simulations were successfully demonstrated that are otherwise not possible with existing single-node resources.
- Research Organization:
- IERUS Technologies, Inc.
- Sponsoring Organization:
- USDOE Office of Enterprise Assessments (EA); USDOE Office of Science (SC), Engineering & Technology. Office of Small Business Innovation Research (SBIR) and Small Business Technology Transfer (STTR) Programs
- Contributing Organization:
- Innovative Computing Lab
- DOE Contract Number:
- SC0020569
- OSTI ID:
- 1733352
- Type / Phase:
- SBIR (Phase I)
- Report Number(s):
- DE-SC0020569
- Country of Publication:
- United States
- Language:
- English
Similar Records
Compare linear-system solver and preconditioner stacks with emphasis on GPU performance and propose phase-2 NGP solver development pathway
Deploy threading in Nalu solver stack