Mert Hidayetoglu of University of Illinois at Urbana Champain presented on his PhD studies.
Talk was entitled “SUPERCOMPUTING FOR FULL-WAVE TOMOGRAPHIC IMAGE RECONSTRUCTION IN NEAR-REAL TIME.”
The abstract of the talk is:
Full-wave methods incorporate all wave phenomena into the image reconstructions by solving the Helmholtz equation with no fundamental approximation. These phenomena include refraction, absorption, diffraction, and multiple-scattering of propagating waves. Although full-wave image reconstruction has promising features and can lead to new imaging technologies, it has been regarded as impractical due to its high computational burden. This talk will be on our effort of making full-wave imaging attainable with two approaches: fast algorithms and supercomputing.
We formulate image reconstruction as inverse multiple-scattering problem which is solved by nonlinear optimization method. The mathematically-exact functional derivative is found with the distorted-Born approximation. However, a large image reconstruction requires solving hundreds of thousands of forward-scattering problems, i.e., inversion of large _N_-by-_N_ dense matrices. As a remedy, we employ the multilevel fast multipole algorithm for solving these forward problems with O(_N_) computational complexity. Furthermore, we use NCSA’s Blue Waters supercomputing facility with CPU+GPU node
architecture for massively-parallel reconstructions. For efficient implementation, we seek low-level GPU optimizations and effective heterogeneous computing.
This talk provides an overview of computational methods, efficient parallelization strategies on large supercomputers, some GPU optimizations, and performance results. Results show good scaling up to 4,096 GPU nodes which provides the largest full-wave image reconstructions to date in near-real time. Several real-life scenarios will be provided where the proposed methodology is especially useful and outperforms conventional approaches in terms of image quality.
Shirt bio of Mert Hidayetoglu:
Mert Hidayetoglu is a Ph.D. candidate ECE Illinois, where he works with
Prof. Weng Cho Chew and Prof. Wen-Mei Hwu. Mert’s research interests
include electromagnetics, fast algorithms for integral-equation methods,
inverse scattering and imaging, and parallel & high-performance
computing. He is among the first two authors of more than a dozen of
conference papers and of several journal papers. He was awarded with the
professor Kung Chie Yeh and Dan Vivoli endowed fellowships by the ECE
department, and with computational science and engineering fellowship by
the college of engineering at Illinois. Recently, he won the second
place at the 2018 IEEE IPDPS Ph.D. Forum poster competition by popular
vote. In Summer 2018, he was a Givens fellow at Argonne National
Laboratory.