Session: Government Agency Student Posters

Paper Number: 173383

Continental-Scale, High-Order, High-Spatial-Resolution, Ice Flow Modeling Based on Graphics Processing Units (Gpus) 

The global mean sea level is rising at an average of 4 mm/year, posing a significant threat to coastal communities and global ecosystems. Approximately 22-500 million people will be affected by annual flooding in 2100 without adaptation. Ice discharge from the Antarctic ice sheet significantly contributes to sea-level rise; however, its dynamic response to climate change remains a fundamental uncertainty in sea-level projections. Numerical ice-sheet models capable of accurately capturing grounding line migration and other important processes are essential tools for these projections. Accurately capturing grounding line migration or the transition between grounded ice (ice sheet) and floating ice (ice shelf) requires spatial resolutions <1km over the area affected by the migration⁠. Conventional ice sheet models may not deliver sufficient computational performance at the continental scale and high spatial resolutions because these models primarily use central processing units (CPUs), which lack massive parallelism capabilities and feature limited peak memory bandwidth. Solving time-independent stress balance equations to predict ice velocity or flow is the most computationally expensive part of ice-sheet simulations in terms of computer memory and execution time. Several recent studies have used stress balance models with complexities lower than the 3-D Blatter-Pattyn higher-order model and spatial resolutions equal to or greater than 1 km near grounding lines to keep computational resources manageable when running simulation ensembles forward in time at the continental scale. These studies partially assess the Antarctic sea-level contribution sensitivity to uncertainties in climate forcing parameterization. Graphics processing units (GPUs) are ideally suited for high-spatial-resolution models since thousands of threads or parallel workers can perform calculations at every grid point concurrently. Furthermore, leveraging GPUs to alleviate the high computational costs associated with ice flow simulations and coupling them with ice thickness and temperature simulations executed on CPUs can provide an enhanced balance between speed and predictive performance. Here, we develop a CUDA C GPU implementation for the Ice-sheet and Sea-level System Model’s (ISSM’s) ice flow solver to alleviate the high computational costs associated with it, and compare it with a standard CPU implementation.  We justify the GPU implementation by applying the price-to-performance metric for up to a ten-million grid point spatial resolution. The methods developed will enable the ice sheet community to quantify the uncertainty bounds in projections with increased confidence, better identify the sources most responsible for the uncertainties in projections, and determine the types of satellite measurements that must be made to reduce uncertainty in projections. Furthermore, the methods and software developed can be extended to accelerate other large-scale Navier-Stokes or incompressible fluid flow applications.

Presenting Author: Ricardo Espin University of North Dakota

Presenting Author Biography: Ricardo Espin is a Ph.D. student in Mechanical Engineering and a Graduate Research and Teaching Assistant at the University of North Dakota. His research centers on high-order, high-resolution ice sheet modeling using the Ice Sheet and Sea-level System Model (ISSM), with a focus on leveraging GPU-accelerated computation for continental-scale simulations. Ricardo is actively involved in academic and engineering communities, including UND Formula SAE, and the Advanced Rocketry Club. He previously served as Vice President of UND Nodak X LATAM, promoting academic and cultural support for Latinx and Spanish-speaking students.

Authors:

Kenneth Mosley University of North Dakota
Ricardo Espin University of North Dakota
Anjali Sandip University of North Dakota