CIS CDUX Researchers Win Best Paper Awards at ICCS 2021 and EGPGV 2021

CIS alumnus Dr. Sudhanshu Sane and researchers from CIS research group on Computing and Data Understanding at eXtreme Scale (CDUX)  have received two Best Paper Awards for his dissertation research on Lagrangian flow. The awards came from two different conferences, both held on the week of June 14-18, 2021. Sudhanshu defended his Ph.D. in May 2020, and the results from both works were in his dissertation, although still unpublished at the time. The first Best Paper Award was at the 21st International Conference on Computational Science (ICCS 2021) for his work investigating Lagrangian representations for data analysis and visualization of cosmology and seismology applications on supercomputers. The research work was selected as the best main track paper from 650 submissions. The second Best Paper Award was at the Eurographics Symposium on Parallel Graphics and Visualization (EGPGV 2021), for his work on accelerating Lagrangian calculations on supercomputers via a communication-free approach. Both awards come with invitations to write extended articles at prestigious journals. Sudhanshu’s Ph.D. advisor, Prof. Hank Childs, was a co-author on both works, and current UO Ph.D. student Abhishek Yenpure was also a co-author on the EGPGV work. More details about the works are below.

Sane S., Johnson C.R., Childs H. “Investigating In Situ Reduction via Lagrangian Representations for Cosmology and Seismology Applications.” (ICCS 2021) [Best Main Track Paper Award]

Computational fluid dynamics is employed across a broad range of applications to study various phenomena of interest. However, due to storage limitations on modern supercomputers, a fraction of the total vector field data generated by a simulation can be analyzed and visualized by scientists. Further, using traditional techniques to perform analysis of time-varying data accurately is challenging due to the under resolved nature of stored data. To address these challenges, UO researchers investigated a novel paradigm, the use of Lagrangian representations of fluid dynamics, for analysis and visualization of two previously unexplored applications: cosmology and seismology. The research involved extracting Lagrangian representations using in situ processing, i.e., coupled with the simulation code, on the Summit supercomputer at ORNL. The research demonstrated the efficacy of Lagrangian representations by measuring in situ encumbrance, presenting a statistical analysis across a range of spatiotemporal configurations as well as a qualitative evaluation of visualizations. The study showed that time-varying vector fields for these applications can be reduced to less than 1% of the total data generated using Lagrangian representations while maintaining accurate reconstruction and incurring less than a 10% overhead to the total simulation execution time.

Sane S., Yenpure A., Bujack R., Larsen M., Moreland K., Garth C., Johnson C.R., and Childs H. "Scalable In Situ Computation of Lagrangian Representations via Local Flow Maps." (EGPGV 2021) [Best Full Paper Award]

Lagrangian representations of flow fields produced by scientific simulations offer significantly improved data storage and accuracy propositions for scientists compared to traditional methods. A challenge with respect to adoption is the increasing computational cost as the scale of the simulation increases. In this study, UO researchers in collaboration with researchers from multiple institutions proposed and thoroughly evaluated a simple, yet novel, model for generating Lagrangian representations. Although there are several works in the area of addressing the scalability of computing Lagrangian representations, the applicability of these studies is limited in an in situ processing context, i.e., an environment with multiple constraints and limited resources. The research work proposes improving scalability by restricting computation to local regions and eliminating synchronization between compute nodes in a distributed memory setting. The study demonstrated that adopting such an inherently scalable solution offers reduced computational costs while maintaining high accuracy analysis in several practical configurations. As part of the experiments, UO researchers utilized as many as 2048 GPUs across 512 compute nodes on the Summit supercomputer at ORNL.

For more information regarding the software libraries employed for these studies:

Alpine Ascent (in situ infrastructure): https://github.com/Alpine-DAV/ascent

VTK-m (portable performance scientific visualization library): https://github.com/Kitware/VTK-m

For more information and details of the research publications:

ICCS 2021: https://link.springer.com/chapter/10.1007/978-3-030-77961-0_36

EGPGV 2021: https://diglib.eg.org/handle/10.2312/pgv20211040