Parallel implementation of the sparse-matrix/canonical grid method for theanalysis of two-dimensional random rough surfaces (three-dimensional scattering problem) on a Beowulf system
Sq. Li et al., Parallel implementation of the sparse-matrix/canonical grid method for theanalysis of two-dimensional random rough surfaces (three-dimensional scattering problem) on a Beowulf system, IEEE GEOSCI, 38(4), 2000, pp. 1600-1608
Citations number
24
Categorie Soggetti
Eletrical & Eletronics Engineeing
Journal title
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
Wave scattering from two-dimensional (2-D) random rough surfaces [three-dim
ensional (3-D) scattering problem] has been previously analyzed using the s
parse-matrix/canonical grid (SM/CG) method. The computational complexity an
d memory requirement of the SM/CG method are O(N log N) per iteration and O
(N), respectively, where N is the number of surface un knowns, Furthermore,
the SM/CG method is FFT based, which facilitates the implementation on par
allel processors. In this paper, we present a cost-effective solution by im
plementing the SM/CG method on a Beowulf system consisting of PC's (process
ors) connected by a 100 Base TX Ethernet switch. The workloads of computing
the sparse-matrix-vector multiplication corresponding to the near interact
ions and the fast Fourier transform (FFT) operations corresponding to the f
ar interactions in the SM/CG method can be easily distributed among all the
processors. Both perfectly conducting and lossy dielectric surfaces of Gau
ssian spectrum and ocean spectrum are analyzed thereafter. When possible, s
peedup factors against a single processor are given. It is shown that the S
M/CG method for a single realization of rough surface scattering can be eff
icently adapted for parallel implementation. The largest number of surface
unknowns solved in this paper is over 1.5 million. On the other hand, a pro
blem of 131 072 surface unknowns for a PEC random rough surface of 1024 squ
are wavelengths only requires a CPU time of less than 20 min. We demonstrat
e that analysis of a large-scale 2-D random rough surface feasible for a si
ngle realization and for one incident angle is possible using the low-cost
Beowulf system.