Better Tools for Supercomputer Research

The U.S. Department of Energy will fund research into a novel approach to improving efficiency of next-generation supercomputer simulations with an award to Rensselaer Polytechnic Institute doctoral candidate Caitlin Joann Ross.

Ross, a researcher in the lab of Chris Carothers—a Rensselaer professor and director of the Center for Computational Innovations—has won a prestigious six-month Office of Science Graduate Student Research Program award to pursue her research at the Argonne National Laboratory in Chicago. Ross is one of 60 awardees chosen for the program, which is sponsored to prepare graduate students for science, technology, engineering, or mathematics (STEM) careers critically important to the DoE Office of Science mission. The Office of Science is the single largest supporter of basic research in the physical sciences in the United States

“Faster, more efficient supercomputers make it possible for us to extract knowledge and insights from ever-larger data sets and address complex questions through simulation,” said Curt Breneman, dean of the School of Science. “Caitlin proposes an ingenious approach to improving the techniques that are pushing the boundaries of supercomputer power, and we are proud to see her early work recognized by the Department of Energy.”

Carothers’ research seeks to improve the design of future supercomputer systems, and much of that work is done through modeling and simulation of supercomputer designs. That in turn leads to a need for more efficient simulations, and in her own work, Ross uses the example of network traffic speed in supercomputer simulators—an aspect of computer speed that has lagged behind the progress of ever-faster processors—to develop tools to improve simulations. One approach to improving network speed is to ensure that network traffic is evenly distributed among network routers, eliminating “hot spots” of network traffic and corresponding inefficiencies.

As part of her doctoral research, Ross has developed a layer in the simulator that collects data about multiple parameters of the simulation, including network traffic. She uses this data to understand the strains that models place on the simulators, and the interplay between the computer and the simulator it is running. Ordinarily, this data must be extracted from the supercomputer for analysis, but in her Department of Energy sponsored research, Ross will develop tools for analyzing the data on-site, potentially extending those tools to offer the ability to alter the simulations as they are running.

“This approach opens a lot of possibilities,” said Ross. “If we can bring the work to the data, we will be able to do the analysis while the simulator is running on some supercomputer. And then we won’t even have to necessarily finish the simulation—we can see what’s going on as its running and maybe even change parameters and improve performance during the simulation.”