- Stampede processes nearly 10 quadrillion operations per second, ranking it among the top 10-fastest supercomputers in the world
- The technology is being used to analyze medical records to develop an early-warning system for deadly brain hemorrhages
- Stampede is also used for finance research, modeling how investors might behave when faced with different systems of taxation
One quadrillion is a number so large, most people can’t conceive what it represents. It’s a one with 15 zeroes after it. It’s one million-billion.
But while a quadrillion may be difficult to conceptualize and even harder to count, Stampede — the world’s 7th-fastest supercomputer, which resides at UT’s Texas Advanced Computing Center (TACC) — processes nearly 10 quadrillion operations per second. That’s the equivalent of 10 petaflops, a measurement of extraordinary speed. For comparison, that’s 40,000 times faster than the best business-class Mac Pro.
And for most people, having a less-than-supercomputer is just fine. If you can play your favorite MMORPG or watch YouTube without it skipping frames every few seconds, your system is probably sufficient.
But for researchers at the McCombs School of Business, even the best desktop is no match for the massive amounts of data and complex models required to tackle some of our most pressing issues.
Predicting Brain Hemorrhages
James Scott, assistant professor in the department of Information, Risk, and Operations Management (IROM), is using TACC’s high-performance computers to study the days that lead up to a subarachnoid hemorrhage, a potentially fatal brain bleed.
Scott explains: “Someone experiences a thunderclap headache — it’s the worst headache they’ve ever had in their life — and it’s because in a particular part of their brain, there’s a burst blood vessel. Most people die from this, and those who do survive can suffer cognitive impairment, so it’s very, very serious.”
For many people, a subarachnoid hemorrhage is preceded weeks beforehand by a sentinel, or warning, migraine. And that migraine is painful enough to send the headache sufferer to the doctor or ER for help. In such cases, if doctors can better recognize when a headache is a precursor to a potentially fatal hemorrhage, they can intervene in time.
For Scott, the answer may be found deep within thousands of anonymized medical records. By analyzing the insurance codes of patients who’ve suffered a subarachnoid hemorrhage, Scott hopes to uncover patterns in demographics, comorbidities, and perhaps even a person’s frequency of doctor appointments leading up to the burst vessel. Finding correlations is what could help doctors identify which patients’ headaches are harmless, and which are harbingers of something worse.
But doing so requires massive computing power. “The data might fit on the hard drive of my desktop, but it would never, ever fit into the working memory,” he says. And that means his desktop can’t do the calculations and modeling necessary to make use of the records.
Scott constantly makes small refinements to his models — a slow and laborious process — until they “pass the smell test” and deliver useable results. What would take hours or days on a regular computer takes a mere 10 minutes on Stampede at TACC, and that efficiency, over time, is crucial.
“TACC really is a game-changer in terms of statistical research,” says Scott.
Streamlining Patient Scheduling
IROM Associate Professor Kumar Muthuraman and doctoral student Ester Wang have also been using TACC’s resources for healthcare research. They want to improve how patients are scheduled for appointments so that clinics are more profitable and patients spend less time in the waiting room.
To do this, Muthuraman and Wang use a patient’s no-show probability along with a clinic’s intake history to predict which kinds of patients will most likely need to be seen during a given timeframe.
Traditional scheduling models have no way of knowing which patients are likely to show up and which patients aren’t, so most clinics overbook across the board to compensate. At the same time, patients are generally given the next available appointment slot without factoring in other patients who may have more urgent issues or be better served with that specific time.
Their model accounts for both: It attempts to minimize wasted time for both physicians, who may be staring at their office ceilings, and patients who are doing the same in waiting rooms. The problem, he says, is that these things don’t happen at the same time.
“All the algorithm is trying to do is to move things around such that things match,” explains Muthuraman.
Wang says that their model must also accommodate all the possible patient flows from one clinic area to another — say, from x-ray imaging to the operating room.
Muthuraman says scheduling patients really well is difficult. “This person is calling for an appointment, but what will tomorrow look like? What will the day after tomorrow look like?” When listing every imaginable eventuality, you find billions and trillions of possibilities, he says.
That kind of massively complex modeling requires the high-powered computing available at TACC. Without it, trying to get this kind of model to give you even one appointment time could take years.
Once Muthuraman and Wang create a working model using Stampede, they’ll develop a smaller-scale program that maintains the accuracy of the original, but that can run on any business computer. Maybe even the one at your own doctor’s office.
Understanding Investor Behavior
Stock market investors have plenty of options when it comes to deciding what to do with their money. That’s where TACC comes in for Associate Professor Stathis Tompaidis.
Tompaidis models a variety of asset allocation options available to investors. That can become quite difficult. “The quantification of uncertainty is very complex, and you need to simultaneously consider many possibilities,” he says.
Taxes are one case he studies: Tompaidis considers how investors might rationally behave when faced with different systems of taxation. Investors might shift their wealth in response to a change in the capital gains rate, for example, versus if the capital gains rate remained unchanged. Because every choice an investor makes today has an effect on what options they will have in the future, modeling possible outcomes can become extremely challenging.
Tompaidis also looks at how age and economic uncertainty influence investors’ decisions. He explains that as investors grow older and earn more money, the economic uncertainty they face can influence the amount of risk they are willing to take in the financial markets.
To model various investment scenarios, Tompaidis will start by creating a simpler model. Once he has an idea of the results and confidence in the model, he’ll turn to TACC to expand the model in order to solve much larger, more realistic problems.
TACC also offers the benefit of speed. Usually, the problems Tompaidis addresses take around 24 hours to solve using TACC; otherwise, he estimates these problems would take 100 times as long to solve. The size of problem Tompaidis can solve depends on how many TACC computers he can use simultaneously. “You can think of the speed, you can think of the size,” he says. “It’s a flip-side of a coin.”
Tompaidis is a long-time fan of TACC, having used their systems over a decade: “Every two or three years they get a better and bigger system, and I move to it.”
Anticipating Consumer Needs
For Associate Professor Frenkel Ter Hofstede and Assistant Professor Jason Duan, marketing research no longer centers around a clipboard and questionnaire. Instead, they now have the buying history of hundreds of thousands of consumers and, just as important, they have Stampede.
They are researching something most shoppers have experienced countless times: What do you do when the item you planned to purchase is out of stock?
Each time a favorite item is out of stock, consumers are forced to make a choice: either to delay the purchase or try something else. Ter Hofstede and Duan are trying to map these behavior changes and uncover how the shopper’s brain learns and adapts to available items.
For example, when your favorite brand of paper towels was sold out and you tried a new brand, was that new product good enough compared to your old standby to warrant switching entirely? What about the second time it was out of stock?
“The whole idea is that out-of-stocks have always been viewed as something bad. But what we’re saying is, it might actually be profitable for the retailer to have an item out of stock,” Ter Hofstede says. That’s because you might opt to buy a store brand in lieu of what you normally purchase, resulting in more money for the retailer.
He adds that in most cases, it benefits consumers, too. Retailers may provide consumers with promotions to try the store brand, which saves them money, and the increased competition from new products in the marketplace helps keep overall pricing low.
But creating a model that can learn from each purchasing variable — what you bought and, in fact, why you bought it — is no easy feat.
“With a regular computer,” Duan says, “you’re constrained. You can only run a few tasks at a time.” For them, that means analyzing each individual shopper’s information separately. With hundreds of thousands of files, it would take years to process that kind of data on a desktop.
Big Computing is Big Business
“That’s where TACC comes in,” says Niall Gaffney, director of Data Intensive Computing. Stampede’s size means a researcher can run many sets of data simultaneously in what’s known as parallel computing.
Stampede’s 6,400 nodes means it’s like working on thousands of connected computers.
The world’s fastest supercomputers are often associated with projects like universe modeling or hurricane forecasting — large-scale scientific endeavors. But business research has made its entrance.
“A lot of what’s coming in the big data space right now is coming from business, and we’re learning as fast as we can from them. It’s the first time that’s happened,” says Gaffney. Big data isn’t new to business, but the complex modeling that’s required to make sense of it is.
TACC’s resources aren’t allocated solely for UT faculty — in fact, 90 percent of Stampede’s time is reserved for the national scientific community — but as Gaffney explains, that leads to collaboration and a sharing of best practices among the country’s top researchers and TACC’s own resident experts.
“It’s a nice partnership going on between McCombs and TACC to bring all these tools and techniques together,” says Gaffney.
— By Adrienne Dawson and Jeremy Simon