Alex Richter, a research economist from the Federal Reserve Bank of Dallas, seeks to unravel the non-linear impacts of the business cycle and monetary policy.
His research requires advanced computing to solve complicated mathematical and statistical problems. For several years, he used the modest-sized high performance computing cluster operated by the Federal Reserve Bank of Kansas City, the Ganymede cluster at the University of Texas at Dallas, as well as Stampede1 and Stampede2 at the Texas Advanced Computing Center (TACC). But he found himself needing more compute time. Moreover, he suspected there were many other research economists in the Federal Reserve System who could benefit additional computing resources.
Knowing that TACC — 200 miles south in Austin — operated several of the world’s largest supercomputers for open science research, in July 2017, Richter travelled to the center to see if he and his colleagues could gain greater access to the supercomputers there.
“I originally went down there thinking, since we’re in Dallas, and TACC is in Austin, maybe there’s a way that we could have some sort of partnership where we could get dedicated access to use Stampede,” Richter said.
He met with Dan Stanzione, TACC’s executive director, and, during a tour of the data center, noticed that several racks that had previously been part of Stampede1 were unplugged.
The system — an Intel/Dell supercomputer with a peak speed of nearly 10 petaflops that debuted in 2012 as the seventh most powerful machine in the world — had recently been decommissioned, he learned. Twenty racks were available for donation. Would Richter be interested in taking them?
Richter was intrigued, but the idea was not without its challenges. For one, the Federal Reserve of Dallas did not have the IT support needed to set up and run the machines, nor a data center to host it.
However, Richter was undeterred. The Federal Reserve accepted the donated equipment, and he spearheaded an effort to get support for a proof-of-concept experiment. TACC would provide the hardware; Chris Simmons, UT Dallas head of research computing, would provide user support and assistance in standing up the machine; the Dallas Fed would establish a hosting environment. The end result of this collaboration would allow Federal Reserve economists from across the country to be able to use the system.
The Dallas Fed rented out space in a local datacenter, purchased two new servers to serve as a master login node and a temporary storage system, and set about creating BigTex, a supercomputer for Federal Reserve economists.
Economists Compute
Previously, the Federal Reserve System had operated several high performance computing environments that contained, in total, about 5,400 compute cores. BigTex, which came online in July 2019, added 15,000 cores — or about 3 times the capacity.
The Federal Reserve is the largest employer of research economists in the U.S. with 400 PhD-level staff. In less than a year, eight of the 13 Federal Reserve banks have signed up to use BigTex and 60 of the 400 economists have used the system to date.
Some researchers, including Richter, are using BigTex to model the non-linear impacts of policy decisions or shocks to the system. In the past, economic models either assumed simplified, linear effects to changes in interest rates or the state of the economy. But with BigTex, researchers are able to tackle more realistic scenarios.
A recent paper from Richter and his collaborators forthcoming in the Journal of Monetary Economics explores various algorithms for estimating non-linear models in cases where the short-term nominal interest rate sinks to zero, creating a ‘kink’ in most models.
“The study asks how well nonlinear solution and estimation techniques compare to linear and quasi-linear methods” Richter said. “It was our most numerically intensive project to date.”
A team from the Federal Reserve of New York, led by Marco Del Negro, is developing algorithms that allow one to quickly update previously trained and tested estimators using new data. This allows economists to rapidly determine what an outcome of a new economic decision could be based on the latest information, without having to re-analyze decades of data. These algorithms make heavy use of parallel computing and hence of BigTex.
They described their results in the Federal Reserve Bank of New York Staff Reports, August 2019.
“BigTex was a game-changer for us,” Del Negro said. “Without it we could have never finished our project.”
Serdar Birinci, an economist with the Federal Reserve Bank of St. Louis, has been exploring the best possible design of unemployment insurance (UI) payments during recessions and expansions. A more generous UI system mitigates the negative effects of job loss, but at the same time incentives staying unemployed. Using BigTex, he found that the best design of the system features more generous payment amounts and much longer payment durations in recessions, as in European policies.
“Analyzing the best design of UI system requires solving a complex economic model under different values of UI replacement rates and UI payment durations,” Birinci said. “A joint determination of these two policy instruments requires solving the model thousands of times. Without having access to BigTex, this would be almost impossible.”
“My colleagues are really happy with BigTex,” Richter said. “It opens the door to research that was previously not able to be done. It’s more efficient and people are benefitting.”
The burgeoning relationship with UT Dallas, TACC, and data center operators is another success. “The fact that we built these partnerships is a very big deal,” he said. “We took an initiative at the local level and turned it into something that has benefited the organization at the national level.”
Not only does the research help economists at the Federal Reserve — it serves as a model for economics researchers in academia and industry.
“TACC produces pie charts of who’s using their resources. One thing that stands out… economists aren’t in them,” Richter said. “It’s important to get more economists and social scientists into the advanced computing fold. I think that’s something worth continuing.”
Header image: BigTex at Flexential Plano data center.
About the Author
Aaron Dubrow is a Science And Technology Writer with the Communications, Media & Design Group at the Texas Advanced Computing Center.