Abstract: Synchrotron light sources are advanced scientific facilities that enable imaging of materials with extremely high spatial and temporal resolutions, helping tens of thousands of researchers to carry out their groundbreaking research every year. Depending on the experimental setup, X-ray experiments at synchrotrons can generate massive amounts of scientific datasets. Iterative reconstruction algorithms are often the preferred method for recovering high-quality 2D/3D images from these datasets, however, their use has been limited to small/medium datasets due to their computational requirements.

In this talk, I will present high-performance optimization methods that we developed for tomographic and ptychographic iterative reconstruction techniques. These will include compute- and memory-centric parallelization techniques with a special focus on communication and data access pattern optimizations on many-core and heterogeneous architectures. I will also present our workflow system that provides an end-to-end X-ray image analysis pipeline between APS and ALCF. I will show the evaluation of our optimizations and workflow system with real-world medium/large scale experimental datasets.