Register now After registration you will be able to apply for this opportunity online.

GPU implementation of an ice flow model

Implement and benchmark a reduced-complexity ice flow model to run in parallel on graphical processing units (GPUs) using the the Julia language and the finite-difference method.

Keywords: glaciology, ice flow model, numerical modelling, GPU, Julia language

Ice-covered surfaces such as ice caps and glaciers provide a visible and significant response to the current acceleration of the global climate modifications. Increasing mean air temperatures and higher variability in precipitation are thus continuously reshaping the glaciers. The temporal evolution of the glaciers is a powerful proxy to better understand the past climate and may be employed to predict or constrain the future changes.
Various mathematical formulations exist to describe the flow of ice over a topographic bed. Reduced-complexity models are often used to approximate ice flow on larger scales as the approximation they build upon allow for lighter code resulting in faster execution times. The depth-integrated-viscosity approximation (DIVA) is a higher-order reduced-complexity ice flow model transforming a 3D ice flow calculation to a 2D map-view model via vertical integration while still considering part of the stress tensor.
In this project we seek for a motivated person to implement the DIVA model to run in parallel on graphical processing units (GPUs). The implementation will be done using the finite-difference method and the Julia language. We provide some core parallel and GPU packages that enable high-performance stencil computations natively on Nvidia CUDA-capable GPUs.
The final codes will serve as benchmark and augment the application examples of the ParallelStencil.jl and ImplicitGlobalGrid.jl modules developed jointly between ETHZ (Ludovic Räss) and CSCS (Samuel Omlin).

Ice-covered surfaces such as ice caps and glaciers provide a visible and significant response to the current acceleration of the global climate modifications. Increasing mean air temperatures and higher variability in precipitation are thus continuously reshaping the glaciers. The temporal evolution of the glaciers is a powerful proxy to better understand the past climate and may be employed to predict or constrain the future changes.

Various mathematical formulations exist to describe the flow of ice over a topographic bed. Reduced-complexity models are often used to approximate ice flow on larger scales as the approximation they build upon allow for lighter code resulting in faster execution times. The depth-integrated-viscosity approximation (DIVA) is a higher-order reduced-complexity ice flow model transforming a 3D ice flow calculation to a 2D map-view model via vertical integration while still considering part of the stress tensor.

In this project we seek for a motivated person to implement the DIVA model to run in parallel on graphical processing units (GPUs). The implementation will be done using the finite-difference method and the Julia language. We provide some core parallel and GPU packages that enable high-performance stencil computations natively on Nvidia CUDA-capable GPUs.

The final codes will serve as benchmark and augment the application examples of the ParallelStencil.jl and ImplicitGlobalGrid.jl modules developed jointly between ETHZ (Ludovic Räss) and CSCS (Samuel Omlin).

- write a GPU parallel DIVA code in Julia GPU
- benchmark the DIVA code implementation on synthetic case and real topography
- perform an Alps ice age run and compare results against recent PISM runs

- write a GPU parallel DIVA code in Julia GPU
- benchmark the DIVA code implementation on synthetic case and real topography
- perform an Alps ice age run and compare results against recent PISM runs

For further information please contact Dr. Ludovic Räss (luraess@ethz.ch).

For further information please contact Dr. Ludovic Räss (luraess@ethz.ch).