Professor Robert van de Geijni
ICES core faculty member Robert van de Geijn is the winner of the 2016 Peter O'Donnell Distinguished Researcher Award. The award recognizes an outstanding research record, impressive and sustained contributions to ICES and the CSEM Program, dedication to the CSEM students, and the distinction and reputation these achievements bring to ICES and The University of Texas at Austin.
Van de Geijn, director of the ICES Science of High-Performance Computing Group and a professor in the UT Department of Computer Science, embodies the award through his research on linear algebra library development, which has wide applications in basic research and industry, and by teaching courses. These include courses at the undergraduate and graduate level.
His work also includes a free massive open online course, or MOOC, which he co-teaches with Maggie Myers, his wife and a lecturer in UT’s Department of Statistics and Data Sciences and his spouse. The course is open to anyone with an internet connection, and has had a total of about 80,000 enrollees over the three times it has been offered, with about 2,400 completing the course.
Van de Geijn’s research involves using linear algebra—the core of scientific computing—to understand and improve how software runs on high performance computers.
“We look at the patterns in the algorithms that make up software and study the science, the system behind those algorithms,” van de Geijn said. “High performance is the goal because people would be upset if we ran on the big machines at the Texas Advanced Computing Center at a fraction of their peak performance.”
A focus of the Science of High-Performance Computing Group is structuring linear algebra libraries to choose the algorithms that are best suited for running on a given high-performance computer. These libraries help make other computer programs run faster and use less energy, features that make them useful to industry and research alike. The approach also has the benefit of generating programs that are mathematically proven to be correct, Myers said.
“Just like you mathematically prove a theorem correct, you prove the program correct,” van de Geijn added. “It’s actually the creation of the proof that gives you the program.”
Two libraries—BLIS and libFLAME-are available online for anyone to use. The research has also influenced other libraries, such as the Elementary library developed by former CSEM student Jack Poulson.
The computing industry, including Intel, HP, Texas Instruments, Movidius, and AMD, uses libraries developed by van de Geijn and his research group to improve their own products.
Timothy Mattson, a senior principal engineer at Intel, says that efficient computers depend on having linear algebra libraries that run quickly. To make sure Intel’s libraries are up to speed, they compare them with van de Geijn’s.
“Basically, we come out with a new release of our math libraries, Robert would beat them with his own [libraries], and then our math library group would figure out what he did and update our software,” Mattison said. “It’s been a healthy rivalry over the year.”
Van de Geijn also played an important part in the development of the 48-core Single Chip Cloud computer in 2010, Mattison said.
“I worked with Robert’s team to move one of their linear algebra packages onto the chip. It was one of the first libraries ported to the chip and played an important role in helping us debug the system.
Improving efficiency is good for business. But it’s also vital for computational research. Devin Matthews, a post-doctoral researcher working with van de Geijn and UT chemistry professor John Stanton, is using the BLIS library to develop codes that describe chemical reactions at the quantum level. An important part of that is computing tensor contractions, an operation that is fundamental to calculations of the kinetic and thermodynamic properties of molecules
“It turns out you can take certain pieces of it and substitute it out, one piece here, one piece there, and you go straight to tensor contraction. So that’s a direct and very visible benefit to chemistry there,” Matthews said. “What people were doing before took thousands and thousands of lines of code.”
Matthews recently received two prestigious fellowships for his own work, the 2015 Frederick A. Howes Scholar in Computational Science for the U.S. Department of Energy and the 2015 Arnold O. Beckman Postdoctoral Fellows Award
Abani Patra, a professor at the mechanical and aerospace engineering department of the University at Buffalo, says van de Geijn’s knowledge of how linear algebra relates to simulation performance was vital as a CAM (precursor to CSEM) graduate student working with ICES Director Tinsley Oden in the 1990s. But it was van de Geijn’s outlook on problem solving that Patra says offered the biggest lesson: many small improvements can add up to big improvements in efficiency.
“It’s a lot of little ideas that he pays attention to. It’s like saving a dollar by saving a penny a hundred times,” Patra said. “He pays very careful attention about how the data is pulled around, how it’s operated on. And at the end of the day those things matter.”
Patra also says his relationship with van de Geijn has remained close over the years. But it began because of van de Geijn’s willingness to see Patra as a peer, even as a beginning Ph.D. student at the institute decades ago.
“He’s always working with students. He’s totally disrespectful of titles and hierarchies and completely respectful of people. He treats the graduate students with the same respect as a director of an institute,” Patra said. “And if the grad student gave him a better idea, maybe a little more.”
In 2013 van de Geijn and Myers started bringing linear algebra research to larger audience through their MOOC, called “Linear Algebra—Foundations to Frontiers,” or LAFF. The goal of the course, which has been taught once a year since 2013, is to teach people how the mathematics of linear algebra relates to how algorithms can be programmed.
“We really wanted to teach people how to program for correctness, which is what our research is all about. So I viewed this first MOOC that we created as a step toward that,” van de Geijn said.
Creating the class was a huge undertaking that involved writing a 900-page textbook, filming 270 video lectures. It was an effort that involved the whole family, van de Geijn said.
“We spent Christmas here. It was a family affair and an unreal amount of work,” van de Geijn said.
But the effort paid off. The class has 4.5 out of 5 stars on edX, the online platform where the classes are offered. And the page is full of positive reviews. One in particular sums up the qualities that the Peter O’Donnell Distinguished Research Award seeks to recognize.
“This course was so amazing, excellent motivation, presentation and exercises (to get insights and intuition). The staff was very active in the discussion forum. This course was like a dream, this was a free high quality education. Really guys I don't know how to thank you, you should so proud for what you offer.”