Bilkent University
Department of Computer Engineering


Memory-Centric, Low Complexity Image Reconstruction for the Exascale Era of Computing


Mert Hidayetolu
PhD candidate at University of Illinois

As we are entering exascale era, advancements in high-performance parallel computing solve everyday life problems those were thought unsolvable. In this talk, I will take X-ray computed tomography (XCT) as an example: measurement data produced by high-resolution scans is often noisy and terabyte-scale. Advanced iterative reconstruction techniques are robust against noisy data, but their computational requirements have made them an exception rather than the rule. As a remedy, we propose MemXCT: a novel memory-centric approach that avoids redundant computations at the expense of additional memory complexity. Considering the next wave of exascale computers, MemXCT favors large resources to scale optimally-well on various heterogeneous supercomputer architectures involving several thousand KNLs and GPUs. Along with portable performance optimizations, results demonstrate unprecedentedly-large iterative XCT imaging in near-real time. To conclude the talk, I will mention a few other critical applications from astrophysics, geospatial imaging, autonomous vehicles, and sparse deep neural networks, where massively-parallel computing with efficient fast algorithms enable cutting-edge science to overcome great challenges we are facing today.

Bio: Mert Hidayetoglu is a PhD candidate at University of Illinois at Urbana-Champaign. He is a member of IBM-Illinois Center for Cognitive Computing and Systems Research led by his advisor Wen-mei Hwu, and SRC Center for Application-Driven Architectures. He previously spent his summers at the University of Hong Kong, Barcelona Supercomputing Center, Argonne National Laboratory, and IBM T. J. Watson Research Center. Mert's research interests include principles and applications of parallel processing, fast algorithms, inverse problems, and supercomputing.


DATE: 15 January 2020, Wednesday @ 13:40