UMIACS Expands Data Center to Support Research Growth

Descriptive image for UMIACS Expands Data Center to Support Research Growth

For almost four decades, a powerful data center in the A.V. Williams Building has been an essential part of the University of Maryland Institute for Advanced Computer Studies (UMIACS), supporting the institute’s mission of incentivizing cutting-edge research and staying at the forefront of technological innovation.

As the institute’s research portfolio continues to expand, the infrastructure needed to support it—particularly in compute-heavy areas like artificial intelligence (AI) and machine learning—also needs to grow.

In April, UMIACS data center staff began coordinating a three-month renovation to expand the center’s physical footprint by 50 percent, growing from 2,300 square feet to 3,450 square feet.

The expansion project will provide additional space for computing hardware and will offer better options to ensure that the data center’s power grid and cooling systems can handle any increased workload, says Derek Yarnell, the institute’s director of computing facilities.

Much of the current workload is handled by multiple racks of graphical processing units (GPUs), powerful parallel computing hardware used for gaming, content creation—and in the case of UMIACS—machine learning and neural networks for high performance computing in areas like AI and computer vision.

In the past five years, UMIACS has increased the number of GPUs in its data center by almost tenfold, Yarnell says. This has allowed researchers to perform certain tasks at almost one petaFLOP of speed—equivalent to more than one quadrillion floating-point operations per second.

“If you measure the amount of computational infrastructure in place that is dedicated specifically toward AI and machine learning research, we are the largest on campus by a wide margin,” Yarnell says. “And we believe this is only the beginning.”

UMIACS faculty are constantly seeking new mechanisms to upgrade their computing resources, says Tom Goldstein, a professor of computer science and director of the University of Maryland Center for Machine Learning (CML).

Goldstein advocates for making economy of scale purchases, wherein CML faculty collectively pool portions of their individual research grants to purchase racks, routers, wiring and accessories needed to support multiple GPU clusters.

The data center expansion matches well with that concept.

“We want to be forward-looking in regard to increasing our computing capabilities, particularly those of us requiring high-memory nodes for projects that focus on large language models and other emerging AI research,” Goldstein says.

Other UMIACS faculty also want to improve both the availability and level of computing power available through the institute’s data center.

Furong Huang, an assistant professor of computer science with an appointment in UMIACS, is working with her graduate students to build a foundation model for sequential decision making, a form of generalist AI that holds the capacity to learn and master diverse sets of tasks downstream.

“Imagine an AI that not only learns quickly, but adapts to your daily tasks autonomously,” says Huang, who is a member of CML.

This work requires analyzing massive amounts of data used in the AI system, a task best handled by a large network of multiple GPUs, Huang says.

Supporting these myriad research efforts falls on Yarnell’s team, a dedicated staff of network engineers and high-performance computing experts assisted by more than 20 undergraduates working at the UMIACS Help Desk.

The institute’s technical staff are involved in the computing process from start to finish, Yarnell says. This includes helping faculty spec out their equipment needs to submit in their research proposals; working with UMD procurement to purchase the equipment; installing all the hardware and software when it arrives on campus; and maintaining the equipment over its technical lifespan by providing security patches and updates.

The tech staff is also responsible for managing and securely storing all the data used in UMIACS research. Current storage capabilities top out at around 3.65 petabytes, with that number constantly increasing, says UMIACS network engineer Jisha Jesudass.

There has also been steady growth in the network capabilities for the institute. When UMIACS moved from the A.V. Williams Building to the Brendan Iribe Center for Computer Science and Engineering in 2019, Jesudass’ job responsibilities included designing and maintaining a high-capacity network—miles of wiring and dozens of sophisticated routers—used to transmit large volumes of data both on and off campus.

She emphasized that the upgrades currently underway in the data center include further improving connectivity capabilities to make the workflow of UMIACS projects more efficient.

“Every job or every researcher right now, they want to minimize latency and delays. This new infrastructure will help a lot with that,” Jesudass says.

For Yarnell (pictured left), the renovation and expansion of the data center remains a work in progress and a labor of love. His involvement with the data center started in 1998 as an undergraduate working at the UMIACS Help Desk. Since that time, he’s seen—and helped lead—much of the data center’s growth.

“Our data center has obviously stood the test of time up to this point—it’s like a Ferrari, a fine-tuned machine that has aged very well,” he says. “But we still think we can continue to innovate it, and continue to build it, in a way that will allow the center to support any new opportunities that are yet to come.”

—Reporting on this story by Shaun Chornobroff, UMIACS communications group

The Department welcomes comments, suggestions and corrections.  Send email to editor [-at-] cs [dot] umd [dot] edu.