The direction that exascale supercomputing will need to follow and the continuing value of visual and other non-computational experts in computer visualizations were the focus of the final two plenary sessions at the PEARC22 conference in Boston on July 13.

Jack Dongarra, director of research staff and professor at the Oak Ridge National Laboratory and the University of Tennessee, Knoxville, surveyed the state of the art as reflected in the new generation of exascale computers in his talk, “High-Performance Computing: Where We Are Today and a Look into the Future.” He argued that moving forward will require both new hardware and hardware tailored to the job of minimizing the communications bottlenecks that force these machines to operate at well below their theoretical capacity.

Donna Cox, plenary speaker and professor emeritus of art and design at the National Center for Supercomputing Applications (NCSA) at the University of Illinois Urbana-Champaign, reviewed the history of visual artists’ involvement with visualization, an approach she pioneered, and the value of nontechnical expertise in technology projects in her talk, “Connecting People for Advancing Our Future: Renaissance Teams and the Art of Interdisciplinary Collaboration for Scientific Visualization.”

The Association for Computing Machinery (ACM) Practice and Experience in Advanced Research Computing (PEARC) Conference Series is a community-driven effort built on the successes of the past, with the aim to grow and increase inclusivity by involving additional local, regional, national and international cyberinfrastructure and research computing partners spanning academia, government, and industry. ACM PEARC22, now taking place, is exploring current practice and experience in advanced research computing, including workforce development, training, diversity, applications and software, and systems and software.

Navigating the Shoals of Hardware Limitations

“Simulation has become the third pillar of science,” Dongarra said. “It augments theory and experimentation. Computational science really is driving the way science is performed today.”

But today’s hardware, which is generally based on parallelizing commodity processors of similar design — of the current Linpack Top500 machines, 78 percent employ Intel processors; a further 19 percent use AMD processors — seldom run at anywhere near peak performance, despite GPU acceleration.

“x86 architecture is really driving high-performance computing in a big way,” he said. “I find that really interesting … We have this monoculture here of high-performance computing.”

In addition to other select Top 10 and Top500 machines, Dongarra used as an example of the Department of Energy’s Oak Ridge National Laboratory Frontier system, currently the top Linpack system. While its theoretical peak performance is 1.7 exaflops, Frontier runs Linpack at 1.1 exaflops — only 65 percent of the theoretical value. Much of that loss is due to communication bottlenecks.

He offered numbers for High-Performance Conjugate Gradient (HPCG) Benchmark, a more representative measure of balanced performance than Linpack – which Dongarra helped create. Frontier hasn’t been tested with HPCG yet (or at least the number hasn’t been made public yet); but looking at other Top500 machines, he pointed to an even worse gap between theoretical and actual performance, with the top HPCG-ranked system, Fugaku in Japan, delivering only 3 percent of its peak performance.

“Most of the machines on the Top 10 list are getting very, very small return … that’s an issue,” he said. “And this issue is because of communication.”

Progress through heterogeneity

Dongarra noted that the dominant technology companies – Apple, Samsung, Google, Microsoft, Amazon, and Facebook – are getting around the problem by investing in bespoke architectures that solve specific problems. While academic computing can’t emulate their funding streams, he recommended an approach that focuses on a redesign of the communication interconnects at various points in the computer architecture that could avoid the bottlenecks.

Such “mixed precision” computing could use high-precision processing, then truncate the results for lower precision when moving them into memory or to another process in the computer. Mathematical techniques that use an iterative approach could converge on an accurate answer using a low-precision result as a starting point.

Using such an approach, as found in the HPL-AI benchmark, Frontier achieved an effective performance of nearly 6.9 exaflops – greatly improved compared with its 1.7-exaflops theoretical performance, let alone its 1.1-exaflops Linpack-ranking performance. This approach uses mixed precision to obtain a fully accurate solution much faster than using the standard 64-bit precision algorithm. Dongarra sees “plenty of opportunity” to improve on current systems’ performance, and further HPC development may hinge on such improvement.

In the future, he added, “High-performance computing will have extreme heterogeneity, with custom systems for each important application.” While today’s biggest systems are based CPUs with GPU acceleration, tomorrow’s may augment with machine-learning processors, application-specific integrated circuits, or possibly even neuromorphic or quantum processors.

Renaissance Teams: Harnessing Diverse Skillsets

Beginning in 1985, Cox began championing the involvement of visual artists in helping scientists and engineers design visualizations that are as vivid as they are informative. The approach became standard procedure at NCSA, leading to a series of visualizations that accelerated science and aided in making its value clear to lay – and government – audiences. Her approach, “renaissance teams,” combined domain experts in a way that may be familiar to researchers today, but did so in a more expansive way, even back then.

Cox established a “10 Commandments” of running renaissance teams that hinged on:

• Setting a common goal or problem to be solved
• Observing mutual respect between team members
• Being willing to learn from teammates
• Recognizing members’ intellectual territory
• Optimizing team size
• Periodically checking goals and progress
• Ensuring team members are not over-committed either within or outside the team’s work
• Naming a project leader or coordinator
• Crediting every team member appropriately
• Ensuring each member is rewarded within their fields

The benefits of this approach, she explained, can be concrete. She cited an example of a 2018 solar-weather funding bill whose prospects for passage by Congress were uncertain. A 2016 documentary on solar storms that her group helped create gave the politicians a better understanding of the importance of the work, helping to get the bill passed. Such “expository visualization,” she argued, stems from the sweet spot of overlap between expert discovery and nonexpert understanding and can move mountains for both educating the public and justifying funding.

“It’s not just the science,” she said. “It’s the communication efforts; it’s the social engineering.” In an era of denialism, effective visualization drawing on artistic expertise can be the key to communicating in a way that connects – and persuades.