Solving electrically large EM problems by using out-of-core solver accelerated with multiple graphical processing units

2013 IEEE Antennas and Propagation Society International Symposium (APSURSI)

2013

We outline an algorithm for block-sparse GPU accelerated out-of-core solver. The algorithm takes advantage of the block-sparsity that exists in method of moment analysis of composite electromagnetic systems with multiple materials. The efficiently of the proposed algorithm is demonstrated on the example of Luneburg lenses up to 40 wavelengths diameter. The approach yields speedup of up to 8 times, compared to generalpurpose GPU accelerated out-of-core solver.

“…As a typical example of EM system that leads to blocksparse MoM matrix, we analyze a Luneburg lens [3]. The permittivity variation is modeled with ten layers of equal thickness.…”

Section: Numerical Example: Luneburg Lensmentioning

confidence: 99%

“…Thus, matrix equations of 120 000 unknowns can be solved at quad-core PC computer in one day. Recently, out-ofcore solvers are accelerated using graphical processor units (GPUs), so that problem with half a million unknowns can be solved in 1-2 days [3].…”

Section: Introductionmentioning

confidence: 99%

Block-sparse out-of-core solver accelerated using GPUs for solving MoM problems

Zoric

Olćan

2013 IEEE Antennas and Propagation Society International Symposium (APSURSI)

2013

“…So, for significant acceleration, it is not enough to parallelize the calculation at GPU. It is also necessary to perform storing/ reading of the matrix blocks in parallel on CPU and in parallel with calculation on GPU [20].…”

Section: Parallelization On Gpus With Cud Amentioning

confidence: 99%

“…In the second class, there are techniques that decrease the memory resources and matrix-fill/solution time for given number of unknowns: (a) iterative techniques [9]- [12], and (b) fast multipole method (FMM) and multilevel fast multipole algorithm (MLFMA) [13]- [16]. Finally, in the third class, there are techniques that enable efficient usage of nowadays hardware resources: (a) out-of-core solution of matrix equation [17], (b) parallelization at CPU based on OpenMP [18], and (c) parallelization at GPU based on CUDA [19], [20].…”

Section: Introductionmentioning

confidence: 99%

Full-wave analysis of electrically large structures on desktop PCs

CEM'11 Computational Electromagnetics International Workshop

Tasić

Olćan

et al. 2011

Method of moments applied to surface integral equations combined with higher-order basis functions enables full-wave analysis in frequency domain of complex and relatively large structures. The electrical size of solvable problems can be further extended using different techniques: symmetry of the problem, "smart reduction" of expansion orders, physical optics driven method of moments, iterative methods, multilevel fast multipole algorithm, out-of-core solver, and parallelization at CPU/GPU. Results are presented for: (1) monostatic RCS of cube of side 50A (100A), (2) beam steering of array of 40 by 40 microstrip patch antennas at 9.2 GHz, and (3) beam steering of 4 by 4 patch antennas at 2 GHz (5 GHz), placed on a 19-m long helicopter.

“…Recently GPUs were successfully used for accelerating outof-core matrix solution in case of complex and electrically large EM problems [8], [9]. Using GPUs the matrix solution time is reduced up to 30 times and matrix fill time becomes dominant for much larger number of unknowns than previously.…”

Section: Introductionmentioning

confidence: 99%

Efficient evaluation of MoM matrix elements using CPU and/or GPU

2012 6th European Conference on Antennas and Propagation (EUCAP)

Zoric

2012

Recent development of GPU parallelized, out-of-core matrix solver enables dramatic reduction of matrix solution time, so that matrix fill time in problems solved by Method of Moments (MoM) applied to Surface Integral Equations (SIEs), becomes dominant. The paper proposes an efficient algorithm for GPU accelerated calculation of matrix elements. The numerical experiments show acceleration from 1.5 times to 21 times, depending on order of expansions and level of integral accuracy.