Editor's note: this article was originally posted on Argonne National Laboratory's website. Written by Jim Collins, Argonne National Laboratory
After finishing his freshman year at Rice University, Max Bowman was looking forward to a summer immersed in the vibrant research environment at the U.S. Department of Energy’s (DOE) Argonne National Laboratory. While his summer definitely didn’t turn out as planned, his virtual internship at the Argonne Leadership Computing Facility (ALCF), a DOE Office of Science User Facility, proved to be a memorable and rewarding experience.
“The effort Argonne put into making the virtual internship a reality was very apparent, and I was able to easily connect with my mentors, other staff and other students,” said Bowman, whose project was focused on the development of quantum-computing tools. “I was still able to work effectively and grow as a scientist and individual. It was a terrific opportunity.”
Bowman was one of 33 students, ranging from high school seniors to PhD candidates, who worked at the ALCF this summer. Each year, the facility brings in a new class of summer students, through programs like DOE’s Science Undergraduate Laboratory Internships and Argonne’s Research Aide Appointments, to work alongside staff mentors on real-world research projects that address issues at the forefront of scientific computing.
“With our summer student program, we’re able to give students practical research experience using some of the world’s most advanced computing technologies,” said ALCF Director Michael Papka. “And it gives us an opportunity to work with talented young researchers, and hopefully inspire them to be a part of the next generation of high-performance computing professionals.”
Here’s a brief summary of a few of this year’s student projects.
Enabling Quantum Computer Simulations
In the rapidly developing field of quantum computing, one of the key challenges is designing methods to model a host of noise sources that can reduce the accuracy of quantum computations. This challenge was the focus of Bowman’s work at the ALCF.
“Scientists must establish techniques to mitigate or account for noise when developing quantum algorithms or examining computation results, so being able to accurately characterize these noise models is crucial,” Bowman said.
To help address this issue, Argonne researchers developed a quantum-systems simulator, called QuaC, that enables the scalable simulation of many kinds of open quantum systems, with special support for collections of noisy qubits, to better understand their properties. Bowman’s internship was aimed at making QuaC more user friendly.
“My work entailed creating an easy-to-use Python plugin that allows scientists to classically simulate quantum circuits with physics-based noise in a scalable manner,” he said.
This involved developing a plugin that combined the features of QuaC and Qiskit, a popular quantum information science library primarily developed by IBM. The Python interface is designed to make it easier for researchers and students to use QuaC to simulate noisy quantum computers with specified physical noise parameters.
In addition to developing the plugin, Bowman also had the opportunity to explore the importance of different components of quantum noise models, as well as methods for inferring these models from quantum circuits run on IBM-Q hardware.
But his interaction with Argonne mentors Hal Finkel and Matthew Otten and the broader laboratory community may have had the most significant impact on his research career moving forward.
“My biggest takeaway from the summer is the power of community and mentorship in research,” Bowman said. “This experience confirmed my interest in quantum information sciences, and I plan to pursue a career related to this field. I am interested in using quantum computation to investigate or predict nanoscale properties of materials.”
Analyzing Congestion on Supercomputers
Network congestion is one of the biggest problems facing high-performance computing (HPC) systems today, affecting everything from code performance to system throughput to user experience.
“It should be of no surprise to developers that congestion on ALCF systems can hurt the performance of the applications that run on them,” said Joy Kitson, a recent University of Delaware graduate who began her PhD at the University of Maryland this fall. “Our work seeks to understand how congestion behaves on these systems, with the goal of using that knowledge to mitigate both congestion and its impact on application performance.”
Working with ALCF computer scientist Sudheer Chunduri and University of Maryland professor Abhinav Bhatele, Kitson used operational data from the ALCF’s Theta supercomputer to analyze communications traffic, with a focus on how congestion is distributed in both space and time across the system. Theta’s Cray Aries interconnect is composed of several groups of network routers that all have a direct link to each other.
“We found that the duration of periods of intense system-level congestion varied greatly, with three such periods lasting on the order of days,” Kitson said. “This congestion was generally distributed evenly across the different regions of the system, but we observed the heaviest congestion on the links connecting routers within the same group.”
A poster detailing her work will be presented at the 2020 International Conference for High Performance Computing, Networking, Storage, and Analysis (SC20).
Beyond learning about the technical aspects of HPC system congestion, Kitson also benefitted from having an open-ended project that allowed her to chart her own path to success.
“It’s important to be able to self-direct on projects like this, particularly when working remotely,” she said. “It was ultimately up to me how to focus my efforts and get useful results from the project, and my mentor’s role was more to verify that I was heading in a reasonable direction.”
Applying Machine Learning to Weather Simulations
Mapping the Journeys of ALCF Users
As a graduate student at the Massachusetts Institute of Technology, Dominic Skinner’s research primarily involves applying mathematical modeling to biological systems. This summer, he was searching for an internship that would expand his scientific computing skills.
“I was looking to spend the summer doing research in some capacity, and it seemed like a good point in my PhD to take a break and work on something else entirely for a while,” Skinner said. “What drew me to ALCF specifically was the work being done in scientific machine learning.”
Through a National Science Foundation Mathematical Sciences Graduate Internship, Skinner collaborated with mentor Romit Maulik, the ALCF’s Margaret Butler Fellow, to apply machine learning techniques to meteorological data.
“When numerically simulating the weather, the finest possible grid on a supercomputer still cannot resolve all of the localized details,” Skinner said. “So, approximations are made to the equations which account for the unresolved dynamics, and recently it has been found that tools from machine learning can provide data-driven approximations that work well. I was particularly interested in how we could accelerate the training and deployment of these models.”
Working with real-world climate data sets, Skinner was able to identify certain periods that required the use of a complicated machine learning model to drive approximations, and other periods where they could use a less computationally demanding model to do the job.
“This has the potential to save significant computational resources when applied to the ensemble predictions that are often used in climate modeling,” he said. “I was mostly working with small models as prototypes, but the ideas are meant to be scaled up to much larger models.”
The hands-on experience Skinner gained with machine learning techniques this summer is something he plans to leverage in his future academic research.
“I’ve learned a lot about scientific machine learning and how it can be used in practical problems,” Skinner said. “The field is still relatively new, but it’s already useful and has given me indispensable tools going forward.”
Becoming a new user at a supercomputing facility involves many steps, which include creating accounts, completing user agreements, setting up a project workspace and learning to submit and run jobs. Making this process as simple and smooth as possible is key to a positive user experience.
Fresh off of graduating from Glenbard East High School in Lombard, Ill., Sarah Iovinelli came to the ALCF this summer to document and analyze the onboarding experience for both ALCF users and support staff. Now a freshman studying user experience design at DePaul University, Iovinelli interviewed several facility users and staff members to create journey maps that detailed the process of getting up and running at the ALCF.
“User journey maps track how and when people interact with an organization and how they feel about the experience,” she said. “This project detailed all of the touch points involved in onboarding at the ALCF, giving the facility a resource to look for potential process improvements.”
Serving as her first foray into the world of user experience design, the project helped Iovinelli boost both her skills and confidence prior to beginning studies at DePaul.
“As a new designer, it can be intimidating to dive into the research stage and figure out how to connect it to potential design opportunities,” she said. “But through my experience at the ALCF, researching the ins and outs of user interactions with the facility became one of my favorite parts of the process.”