Science simulation, visualization, data, and learning applications will greatly benefit from the massive computational resources available with future exascale systems. Researchers in the Argonne Leadership Computing Facility’s (ALCF) Aurora Early Science Program (ESP) are blazing the trail toward reaping those benefits from the U.S. Department of Energy’s (DOE) Argonne National Laboratory’s upcoming Aurora exascale supercomputer.

Work by ESP researchers will help to ensure that critical scientific applications are ready for the scale and architecture of the Aurora machine at the time of deployment. There are currently around 250 researchers involved in pre-Aurora ESP research. According to Timothy Williams, Deputy Division Director of Argonne’s Computational Science (CPS) Division and ALCF Co-Manager for the ESP, “As one of the first exascale systems for science in the world, Aurora should deliver siginifcant scientific results, via the Early Science Program.” The ESP is already producing some exciting research and providing insights for system architecture and infrastructure changes slated for the future Aurora supercomputer.

ESP projects represent research so sophisticated that it has outgrown the capability of today’s leadership-class supercomputers—the selected ESP research projects require exascale computational capabilities. Research Principal Investigators (PIs) submit proposals to ALCF describing their research into a specific scientific problem and why it needs to run on an exascale system.

The ESP awards pre-production computing time to research teams working to prepare key applications and software for the Aurora supercomputer. ESP researchers are granted access to hardware and software running on a pre-Aurora configured supercomputer. Argonne’s Theta supercomputer has been extensively used by the ALCF staff and ESP researchers who are preparing for Aurora.

ESP research projects are in the areas of chemistry, physics (high energy physics, fusion energy, cosmology), biosciences (cancer treatment informatics, modeling metastasis, brain connectomics, molecular dynamics of cell membrane transport proteins), engineering (aerodynamics, nuclear reactor coolant, combustion in coal boilers), materials science (functional materials, semi-conductors).

William Tang, professor of astrophysical sciences at Princeton University and principal research physicist with the DOE’s Princeton Plasma Physics Laboratory (PPPL), is leading an ESP project that is one of the more successful efforts in artificial intelligence (AI) for science using pre-exascale systems. His work is focused on using deep learning and exascale computing power to improve the behavior of fusion reactors aiming to produce sustainable clean energy.  Tang’s AI research studies disruptions in confinement devices called tokamaks, which use a powerful magnetic field to confine hot plasma to produce controlled thermonuclear fusion power.

Engineers working with the potential energy source have estimated a window of only 30 milliseconds to control instabilities that can disrupt the energy production process and damage the plasma confinement device. As part of the ESP research, Tang and colleagues use Princeton’s Fusion Recurrent Neural Network (FRNN) code containing convolutional and recurrent neural network components to integrate both spatial and temporal information for predicting disruptions in tokamak plasmas. The hope is to increase warning times and work toward heading off disruptions before they happen—keeping the fusion reactions going and producing sustainable clean energy.

Another of the ALCF’s notable ESP projects is led by Katrin Heitmann, Deputy Division Director in the High Energy Physics Division at ANL. Heitmann and team perform research using computational cosmology to understand the large-scale behavior of the universe. The research seeks to understand fundamental aspects the cosmos such as dark matter, dark energy and to help understand why the universe’s rate of expansion is accelerating.

The cosmology simulations are carried out using the Hardware/Hybrid Accelerated Cosmology Code (HACC) developed at Argonne, based on an early effort at Los Alamos. HACC is the only cosmology code suite designed for extreme-scale simulations regardless of a supercomputing system’s architecture. The team also uses advanced data science techniques in conjunction with observational data. These techniques have been developed in collaboration with statisticians over a period of many years. More recently, AI methods have been trained using a large set of images generated from cosmological simulations run with HACC.

Moving toward exascale requires not only moving applications to new computer architecture, but it also requires:

• Code and workflow development
• Preliminary studies
• Scaling and optimization

The ESP provides resources and support across these requirements to help research teams prepare their applications for the architecture of the new supercomputer.

The ALCF computational scientists work with ESP researchers to help with troubleshooting, coding, optimizations for parallelization and GPU acceleration, getting the ESP research applications to run in the pre-Aurora environment. Members of the ALCF team also provide support for projects with big data, deep learning (DL), or machine learning (ML) requirements. “Each of the computational scientists working with researchers speaks the language of the relevant domain sciences as well as high-performance computing. In most projects, preliminary studies must be done in advance to verify that the planned exascale research campaigns will succeed,” states Williams.

The ALCF provides a variety of Aurora-related training opportunities including hackathons, workshops, dungeon sessions, and webinars. Some focus around developing, porting, optimizing code with the Aurora SDK and early Intel GPU hardware housed at Argonne’s Joint Laboratory for System Evaluation (JLSE).

Williams indicates, “The ALCF Data Science team (headed by Venkat Vishwanath, ALCF Co-Manager for the ESP program) is establishing a data science supercomputing software environment on Theta, which is the closest environment to what we plan to have on Aurora—it includes the Balsam workflow manager, support for optimized Python functionalities, ML/DL frameworks, parts of the Big Data stack—all optimized for HPC and scientific applications.”

The Exascale Computing Project (ECP) is developing an exascale software stack, including software needed by application developers writing parallel applications targeting diverse exascale architectures. ALCF partners with and participates in the ECP to deploy this stack for Aurora. Software is also being developed for large scale and in-situ visualization and analytics projects.

The future Aurora supercomputer will also include the Intel Distributed Asynchronous Object Storage (DAOS) I/O technology, which alleviates bottlenecks involved with data-intensive workloads. DAOS, supported on Intel Optane persistent memory, enables a software-defined object store built for large-scale, distributed Non-Volatile Memory (NVM). The combination of Intel Optane persistent memory and DAOS, recently set a new world record, soaring to the top of the Virtual Institute for I/O IO-500 list. DAOS will be the primary data storage platform for ESP and production science projects on Aurora—a major advance beyond conventional parallel file systems.

Argonne is a key participant in the development of oneAPI, a unified and scalable programming model to harness the power of diverse computing architectures in the era of HPC/AI convergence. The oneAPI initiative – supported by over 30 major companies and research organizations and growing – will define programming for an increasingly AI-infused, multi-architecture world. The oneAPI unified programming model is designed to simplify development across diverse CPU, GPU, FPGA, and AI architectures

“Through Argonne’s deep investment in science projects using data-intensive and machine-learning methods, Aurora will advance the state of the art for complex scientific workflows at large scale—especially those including experimental/observational data. Aurora will play a big role here,” states Williams.

References

• Kates-Harbeck, J., Svyatkovskiy, A., & Tang, W. (2019). Predicting disruptive instabilities in controlled fusion plasmas through deep learning. In Nature (Vol. 568, Issue 7753, pp. 526–531). Nature Publishing Group. https://doi.org/10.1038/s41586-019-1116-4
• Habib, S., Pope, A., Finkel, H., Frontiere, N., Heitmann, K., Daniel, D., Fasel, P., Morozov, V., Zagaris, G., Peterka, T., Vishwanath, V., Lukić, Z., Sehrish, S., & Liao, W. (2016). HACC: Simulating sky surveys on state-of-the-art supercomputing architectures. New Astronomy, 42, 49–65. https://doi.org/https://doi.org/10.1016/j.newast.2015.06.003
• Argonne Leadership Computing Facility – Early Science Program https://www.alcf.anl.gov/science/early-science-program
• P. Straatsma, K. B. Antypas, and T. J. Williams, editors, Exascale Scientific Applications: Scalability and Performance Portability, CRC Press, 2017, ISBN 9781138197541.

Author: Linda Barney is the founder and owner of Barney and Associates, a technical/marketing writing, training, and web design firm in Beaverton, OR.