Scientists create unique tool to help supercomputers run faster

Members of the Analytics and AI Methods at Scale group in the National Center for Computational Sciences at ORNL developed the mixed-precision performance benchmarking tool OpenMxP. From left are group leader Feiyi Wang, technical lead Mike Matheson and research scientist Hao Lu. Credit: Carlos Jones/ORNL, U.S. Dept. of Energy

Researchers have developed a unique tool that helps the world’s most advanced supercomputer, Frontier, perform complex calculations more effectively.

Before we dive into the details, let’s simplify some concepts.

Think of supercomputers as super powerful and super fast computers that can solve problems and do calculations much faster than regular computers.

They are used for various scientific and engineering tasks, like predicting weather, studying diseases, and many more.

Usually, supercomputers use something called double-precision arithmetic, a 64-bit system, which is like using a very accurate ruler to measure something.

But now, there’s a way to use a less precise, but faster measuring tool, known as mixed-precision arithmetic (16 or 32 bits), often calculated by GPUs. This is especially handy for data science and artificial intelligence (AI) tasks.

However, until recently, there wasn’t a widely available tool to test how well supercomputers perform using this less precise, but faster, measuring tool at extreme scales.

Enter OpenMxP

To tackle this, scientists at the Department of Energy’s Oak Ridge National Laboratory developed OpenMxP, a special software, just in time for the launch of Frontier in May 2022.

OpenMxP can test how efficiently supercomputers, like Frontier, perform these mixed-precision calculations. The good news is, they’ve also made OpenMxP available for free to other computing facilities, meaning more supercomputers can benefit from it!

Frontier is important as it’s a leader in computing technology and is fundamental for the advancement of science in the U.S. OpenMxP helps understand how well Frontier, and other similar systems, are working and how they can be improved, revealing how making small changes can lead to big improvements in speed.

Testing and Development

Developing OpenMxP wasn’t a straightforward task. Frontier was still in the making, and the researchers had to start by understanding the problems the tool was supposed to solve and learning from scientists who had worked on similar issues.

Initially, they used another powerful system, Summit, for testing OpenMxP since Frontier wasn’t ready. This way, when Frontier was assembled, OpenMxP was ready to roll and worked efficiently, helping Frontier achieve impressive performance records.

Beyond Benchmarking

But OpenMxP does more than just put a number to how fast supercomputers are. It gives insights into how well these supercomputers are operating and shows the advantages of using systems equipped with GPUs capable of mixed-precision calculations.

Many simulations solve large systems of equations and traditionally, it’s all done with double precision. But with GPUs that can do low-precision calculations faster, a new and attractive solution process has been enabled.

Furthermore, OpenMxP can be used as a tool to solve specific scientific and engineering problems faster and more efficiently than ever before.

In 2022, a team used OpenMxP for a software called TwoFold that predicts how strongly a drug will bind to a pathogen and the 3D structure of how it will attach to the target. TwoFold was even named a finalist for a prestigious prize for high-performance computing-based COVID-19 Research.

Key Takeaways

The development of OpenMxP represents a significant leap in computational science.

It not only serves as a measuring tool to gauge the efficiency of supercomputers but also acts as a solver for specific scientific and engineering problems, potentially accelerating solutions in various fields like medicine and environmental science.

By making it open source, the developers of OpenMxP have ensured that this revolutionary tool is accessible to a broader scientific community, fostering innovation and collaborative advancements in science and technology.

Follow us on Twitter for more articles about this topic.

Source: Oak Ridge National Laboratory.