Turbulence simulations can model what happens in a tornado, a galaxy or the swirls generated from aircraft wing tips. Image courtesy of Karthik Duraisamy.
ConFlux is designed to have simulations and large datasets interface with each other during a run. The nodes will include CPUs, GPUs, large memory and fast interconnections. The storage of the big data will be on a three-petabyte (that’s 1015, or a 1,000 terabyte) hard drive. These nodes are also designed to allow for data-intensive operations.
The HPC system will allow U.S. researchers to focus on computing infrastructure and data-driven physics, and to address studies previously done on supercomputers.
"Big data is typically associated with web analytics, social networks and online advertising. ConFlux will be a unique facility specifically designed for physical modeling using massive volumes of data," said Barzan Mozafari, UM assistant professor of computer science and engineering and overseer of the project.
As previously mentioned, a common problem with many simulations is scale. If you are looking into the material property behaviour at the atomic scale, or problems with complex systems like climate, scale is of primary importance.
Karthik Duraisamy, UM assistant professor of aerospace engineering at UM, notes that many of the world’s powerful computers are able to handle these problems by using approximations. Unfortunately, these approximations are not always accurate enough to answer some of the harder engineering and science questions.
"Such a disparity of scales exists in many problems of interest to scientists and engineers," said Duraisamy. "We need to leverage the availability of past and present data to refine and improve existing models."
To address this problem, experiments, measurements and simulations with limited scope can be used. Algorithms can crunch the data to make the needed predictions. The machine can learn more algorithms as the simulations, experiments and measurements improve and increase in number.
"It will enable a fundamentally new description of material behaviour — guided by theory, but respectful of the cold facts of the data. Wholly new materials that transcend metals, polymers or ceramics can then be designed with applications ranging from tissue replacement to space travel," said Krishna Garikipati, a professor in the departments of mechanical engineering and mathematics at UM.
Some research projects that will be using ConFlux include:
- Combining non-invasive imaging with physical models for blood flow to treat cardiovascular disease.
- More accurate turbulent simulations to predict swirls and eddies. This could improve airplane design, weather forecasting and climate science.
- Studying the effects of climate change on clouds and precipitation.
- Simulations of galaxy formations using galaxy-mapping studies to better understand the role of dark matter.
A Center for Data-Driven Computational Physics will be built to manage the ConFlux HPC system. The project will be funded by a $2.42 million NSF grant and $1.04 million from UM. The project fits in line with President Barack Obama’s National Strategic Computing Initiative to use vast data sets and to increase computing power.
When have you hit a scale problem in your simulation and how did you solve it? Comment below.