HPC Reveals Glacial Flow

Print Friendly, PDF & Email

In this special guest feature from Scientific Computing World, Robert Roe looks at research from the University of Alaska that is using HPC to change the way we look at the movement of ice sheets.

Today’s scientists are much more informed about the potential for climate change and melting of glacial ice.

However, while some effects of climate change are well understood, the flow and changes of glacial ice have traditionally been observed and measured, and researchers lacked the computational resources to model these effects.

University of Alaska researchers at the university’s Geophysical Institute have now created computational models that can accurately predict the movement of glacial ice flow using HPC.

Andy Aschwanden, research assistant professor, Geophysical Institute, University of Alaska Fairbanks (UAF), is studying the ice flow of the Greenland ice sheet. Aschwanden and his colleagues are hoping that this research will help predict the future evolution of the Greenland ice sheet.

The Geophysical Institute consists of seven major research units and supporting facilities, including space physics and aeronomy; atmospheric sciences; snow, ice, and permafrost; seismology; volcanology; remote sensing; and tectonics and sedimentation.

Over the past two decades, professor Aschwanden stressed that large changes in flow have been observed in outlet glaciers and that the rate of melting ice has increased. This has been calculated at a 17 per cent increase in ice-sheet wide melt between 2000 and 2005. While researchers are reasonably confident about the rate of ice melting, more research was needed to predict future changes and their potential impact on glacial ice.

Complex challenges demand HPC

UAF researchers have been developing the open-source Parallel Ice Sheet Model (PISM) since 2006. The PISM model was created to enable researchers to model glacial ice flows accurately. But, due to the complex nature of the simulations, the models can be incredibly computationally intensive.

Ice sheets, which are essentially continent-size glaciers, are constantly moving. Just like other fluids, glaciers flow downhill – driven by gravity. PISM simulates the movement of the ice fluid, and its temperature, just as a weather forecasting model can predict atmospheric conditions.

Ice sheets contain a large amount of frozen water which is currently above sea level. Ice sheets in Greenland and Antarctica for example, are more than two miles thick and sitting on land. If ice sheets flow faster or slower, or the way they flow changes, then this affects the rate at which they can raise sea level, so it is important to understand if climate change is impacting the rate of ice flow.

The simulations needed to track the massive ice sheet’s progress require large resolutions and must process huge amounts of data. This means that access to more powerful computers and faster networking solutions can not only reduce the time to derive scientific insight but also increase the resolution of the model creating a more accurate prediction of glacial ice flow.The research paper covering this project – which was published in the journal, Nature, noted that this project was made possible by the combination of several factors. The first factor is the substantial improvement in scientific knowledge of subglacial topography, particularly in deep channel-feeding outlet glaciers.

This knowledge was slowly built up over several different projects such as the NASA airborne mission Operation IceBridge (OIB) which began in 2009. OIB deciphered several 1,000 km radar-derived ice thickness profiles. The remaining gaps were covered by mass-conserving interpolation methods which were used to derive flow-compatible, high-resolution maps of ice thickness and subglacial topography.

This knowledge, combined with advances in HPC performance and code parallelisation, have made high-resolution ice sheet modelling feasible.

The research notes that combining these advances allows the researchers develop a set of numerical experiments to investigate whether spatially complex flow patterns in outlet glaciers can be captured in whole-ice-sheet simulations using only ice-sheet-wide (spatially uniform) parameters, without local ‘tuning’ applied to individual grid cells.
The study used the PISM model, coupled with models of subglacial hydrology and basal sliding, to simulate the velocity field of the Greenland ice sheet at a resolution of <1 km.

The paper states: ‘We demonstrate that outlet glacier flow can be captured with high fidelity if ice thickness is well constrained and vertical shearing, as well as membrane stresses, are included in the model (without solving the full-stress configuration).’

However, the study also reports that computing flow from vertical shearing alone or using low-resolution ice thickness leads to poor agreement with previous observations.

Overall root mean squared (RMS) velocity differences decrease with increasing model resolution. This indicates that ongoing improvements in the mapping of subglacial topography, together with improvements in modeling resolution, go a long ways towards improved whole-ice-sheet numerical simulations’ the research notes.

Providing the computational muscle

The computational muscle behind this research project comes from the UAF’s Geophysical Institute which houses two HPC systems ‘Chinook’, an Intel based cluster from Penguin Computing and ‘Fish’ a Cray system installed in 2012 based on the Cray XK6m-200 that uses AMD processors.

Chinook, the system used for this project, is known as a Penguin Computing Community Cluster and operates under a ‘community condo model’ that allows principal investigators (PI) to obtain access to computational resources. This includes a regulated datacentre, network connectivity, equipment racks, management and technical staff – resources they may not be able to procure, allowing users to focus time and energy on research, rather than owning and operating individual clusters.

Participants in the condo service then share unused portions or elements of the computational resources they add to Chinook with each other and non-invested users – such as students who might not pay a fee for access. A queue management system gives vested PIs top priority to the share they have purchased whenever the PI needs the resource.

The system operates a number of nodes for the community and then two-tier shareholder nodes, including some nodes dedicated purely to the dedicated use of shareholders.

In January 2017, with support from the M J Murdock Charitable Trust, the Geophysical Institute, UAF vice chancellor of research, UAF International Arctic Research Center, and UAF IDeA Network of Biomedical Research Excellence, UAF Research Computing Systems engineers upgraded their HPC system.

The upgrade doubled the number of available cores to 1892 and increased interconnect speeds from 40 Gb/s up to 100 Gb/s by deploying Mellanox InfiniBand technology across multiple racks of their HPC system.

This community, condo model project launches a significant change in how high-performance computing resources are offered to the UA community,” said Gwendolyn Bryson, manager of Research Computing Systems at the UAF Geophysical Institute. “Chinook is more capable, more flexible, and more efficient than our legacy HPC resources.”

This increase in performance will allow researchers to carry out more complex simulations and to process data much faster as the interconnect upgrade has significantly increased the speed of data transfer. This may allow future glacial studies at a higher resolution to more accurately model the structure and movement of glacial ice flow.

Finding answers

The research project successfully managed to portray Greenland’s flow structure in unprecedented detail. The paper reports that most outlet glaciers considered in the study were captured at grid resolutions of less than 1 km. The results also demonstrate that the spatial variability can explain spatial variability in flow in ice thickness.

The research paper states: “We find that inversion of surface properties for individual glaciers is not essential to reproduce the overall flow pattern. Using simple parametrisations of basal motion and subglacial hydrology, we find good agreement between simulated and observed spatial flow patterns.”

However, the work is not yet over as the researchers note that the reproduction of the velocity magnitude requires further improvements, especially in the transitional zone 100 to 300km inland. The study also lists some areas of possible investigation for future improvements including inadequacies in parameterising basal motion and subglacial hydrology which may account for the disagreement between observed and simulated speeds.

The researchers also list other considerations such as the fact that not all heat sources that affect the viscosity of ice are accounted for in the model. For example, refreezing of surface meltwater (‘cryo-hydrologic warming’) can soften the ice and enhance flow, which may be relevant in the transitional zone.

Ultimately this research will help other researchers come to a better understanding of the potential for sea level rise as ice flows from glaciers into the sea. While these simulations can go some way towards an understanding of glacial flow in Greenland, they still need to be applied to other glaciers before gaining a complete understanding of the implications.

Our simulations have implications for efforts targeted at projecting 21st-century sea level rise,” stated the paper.

It also reported that the Intergovernmental Panel on Climate Change (IPCC) Fourth Assessment Report called for the need to resolve the full stress configuration in ice-sheet models to simulate changes in outlet glaciers.

Since then, ice sheet models have seen substantial improvements in their representation of flow physics.

We find that models that resolve both membrane and vertical stress gradients are capable of reproducing the observed flow structure with high fidelity.

“In regions with large transverse velocity gradients, such as sheer margins, the mismatch between observed and simulated flow may be further reduced by resolving additional components of the stress balance,” the researchers noted.

This story appears here as part of a cross-publishing agreement with Scientific Computing World.

Sign up for our insideHPC Newsletter