Computers—and people—work in parallel at new laboratory
A trio of University of Wisconsin-Madison engineering professors has launched a new high-performance computing laboratory. With a broad range of industrial and research partnerships, the lab will leverage the capabilities of a unique computer cluster and more generally expand access to and applications for supercomputing resources.
The Modeling, Simulation and Visualization Center will combine the expertise of Mechanical Engineering Bernard A. and Frances M. Weideman Professor Vadim Shapiro, Associate Professor Krishnan Suresh and Assistant Professor Dan Negrut. Modeling essentially means reducing a problem, such as the mobility of a vehicle or robot, to a series of equations. Simulation is the process of using a computer to solve these equations, which are too complex to do by hand. Visualization is turning the solution data into usable information via animations or graphs.
The heart of the lab is a computer cluster composed of 5,760 scalar processors split across 24 graphics processing units (GPUs). Negrut, who is the current lab director, compares the GPUs to a group of jugglers. "This is like 5,760 jugglers split into 24 groups of 240," he explains. "For them to work together, you have to have some type of protocol and understand who's passing the sticks—or data—to who, how often and how many," he says.
If the team can establish a protocol to allow all 24 GPUs to work in tandem, the result will be tremendous computational power. The cluster handles very complex problems, such as determining which of one million bodies (such as grains of sand) collide with each other. The cluster calculated this dynamics problem, with 15 million total unknown variables, in around three minutes. The cluster has also solved a collision detection problem with six billion contacts, approximately the number of contacts in 35 cubic feet of sand.
As the team continues to increase the number of GPUs that can be used in parallel (four is the current maximum), Negrut expects that similar complex problems will be solved an order of magnitude faster than can be done with current technologies.
Similar clusters exist at financial institutions, and GPU computing is common in the oil industry and elsewhere, but the UW-Madison lab is unique in its breadth of applications. Several industrial and federal research partners already are interested; National Science Foundation research grants support Shapiro, Suresh, and Negrut, and NVIDIA donated half of the cluster hardware. Microsoft has funded educational initiatives and provided the underlying operating system for the cluster. The U.S. Army; NASA's Jet Propulsion Laboratory; FunctionBay, a simulation company in South Korea; Mihai Anitescu of Argonne National Lab; and Professor Alessandro Tasora from the University of Parma, Italy, are also involved.
This broad range of partnerships expands on the interdisciplinary nature of the three founding professors. Shapiro studies computer-aided design and computational geometry. Suresh focuses on nonlinear finite element analysis, and Negrut researches computational multibody dynamics—a range of complementary topics that each rely on high-performance computing.
"We're leveraging this hardware asset and trying to team up and combine our knowledge to answer challenging problems that people in industry might have," Negrut says.
The lab's outreach extends beyond industry and other universities. Since 2008 Negrut has organized a week-long summer camp, called Promoting the Computational Science Initiative, or ProCSI, for underrepresented high school students interested in computational science, engineering and college in general. Additionally, high school students from a Madison high school have attended lectures at UW-Madison by lab partners, and two students worked at the lab for a month during the summer of 2010.