Monday 23 October 2017

Mercedes handles the competition because it knows how to handle data, too

reader comments 1

AUSTIN, Texas—History happened Sunday at the Circuit of the Americas. Formula 1 driver Lewis Hamilton won for the fifth time in six years at Austin, inching him closer to a fourth world championship this year. And on a macro scale, Hamilton’s victory sealed a fourth straight Formula One constructors’ championship for the Silver Arrows team at Mercedes. According to ESPN, that makes Mercedes the first team to win consecutive championships across a major regulation change.

How does a team achieve such sustained dominance—Mercedes has won a staggering 51 of 59 total races between 2014 and 2016—in an era where the sport has witnessed an infusion of more money, more engineering talent, and more of those aforementioned regulations? If you listen to members of the Mercedes-AMG Petronas Motorsport tech team tell it, the answer starts in the team’s network stacks.

“The winning direction today is understanding what kind of problem are you trying to solve. Engineers are all interested in solving problems, but my mantra for a while has been ‘make sure you’re solving the right problem and not just the first one that comes along,’” Geoff Willis, Mercedes-AMG Petronas Motorsport’s former technical director and the team’s newly minted digital, engineering, and transformation director, tells Ars.

“With the top teams, there’s much less trial and error and more predictive understanding. So before we go to a race like here in Austin, we’ve done weeks and weeks of simulations of how to set the car up; drivers have done simulations in it, too. We have a good picture of what to expect, so what we look for when we get here: ‘Is there anything that alerts us to the car not behaving as we expect?’ If so, then we have a lot of what if studies to rely on.”

The ability to recognize and address reliability issues swiftly was certainly the theme when Ars got the opportunity to tour the Mercedes garage ahead of this weekend’s race. That invitation didn’t come from Mercedes, rather it came from Pure Storage, the California company that partnered with the carmaker early in 2016 to bring flash storage both to the factory and trackside. Network gear may seem like only a small piece of Mercedes’ winning puzzle, but the IT-minded on pit row quickly stressed how important their new storage solution can be.

Pure Storage's teaser video outlining the Mercedes partnership.

Simple logistics

Bottom-line numbers made the switch to Pure Storage flash arrays an easy decision for Mercedes, especially considering that hard disk drives were still in vogue within F1’s last decade. So in a sport where garage size can vary week to week (with Austin being on the smaller end: 2.5 Austin garages would fit in the Abu Dhabi one, according to the team), the new devices save a tremendous amount of space. Matt Harris, Mercedes' head of IT, says the team reduced the size of its networking stacks by nearly 70 percent, enough to make up the device cost with only two years of freight savings. “If you keep the weight down and save on cost, you can invest in other performance areas,” says Christian Dixon, a partnership manager on the Mercedes team. “And the more room we can save, the more equipment we can bring.”

More important than physical logistics improvement, however, the Pure Storage arrays helped Mercedes store and access its whopping amount of data more efficiently. Pure Storage says its technology minimizes the amount of data needed to be stored in a location two times more efficiently than its competitors, and (crucially for motorsport) it can transmit data in real time. As you might expect, the Mercedes team has needs more urgent and much larger than the Exchange archives of your average office space.

“Think of the cars as sensors going around the track, picking up info on acceleration, vibrations, pressures, temperatures—we have over 200 sensors on the car,” Dixon says. “We record over 100 times a second with 1,000 channels of data—we’re creating 1.8 billion data points.”

“And we generate 500GB in a race weekend, not just from the car but from everything we do,” Harris adds. “In fact the processing power of the car is the biggest problem—if the processor was faster, we could get data off faster. But now we have to compromise by weighing speed of offloading, speed of turnaround for the car to make decisions, and how much data we want to generate.” (Harris notes the ECU processor, dating back to 2009, is practically the only thing on the car that hasn’t radically changed in recent years.)

Trackside, Harris says 30 or so teammates are dedicated to looking at the data, and updating their systems from relying on legacy servers to the Pure Storage arrays has enabled those datawatchers to act more quickly. “[With the old system], they knew it’d be one to two minutes to open the file, read through the data, and make a decision,” he says. “Opening the wrong bit of data would add time. Now, Pure brings the process down—you can actually make the wrong decision on which piece of data to open without compromising the next run of the car.”

For a real-world example of this new infrastructure supporting the on-track efforts, Harris points to this year’s race in Singapore. Valtteri Botta, Mercedes’ other world-class driver, kept telling the team he felt a cut in the engine. “But the guys kept saying, ‘No you’re not, you’re not,’” Harris says. “But they had to keep getting more refined on the data to see it; it ended up being a 13,000th of a second and Valtteri could feel it. It was a magnetic field the bridge created.”

The future, where ML meets Mercedes

As you may guess based on their recent history, the Mercedes team is already thinking extensively about where data analysis and storage need to be in the F1 future. To that end, Harris says, the team has started toying with ways to leverage modern machine-learning and artificial intelligence techniques, too. At their factory back in Brackley, England, they rely on Pure Storage Flash Blades (a scalable, parallel storage solution) to store all simulation results and historic data. Mercedes then combines that with another partnership, this one with a company called Tibco that produces software capable of leveraging machine learning for big data analytics.

“We always knew collecting data was a good thing, but we weren’t using it efficiently—it was hard to know what you want to find out and what’s useful to do,” Dixon says.

“So we asked, ‘How can we get rid of the normal data?” Harris continues. “We still keep that on a filer, but we don’t have to waste our time to look at it if it’s normal. What you want is abnormal data—is it abnormal because we made a change, or is there an issue, or is some kind of trend happening? We wanted to start automating the search for some of that since there’s only so many sets of eyes. These machine-learning, deep-learning techniques we’re beginning to look at it—and we are new to it, though learning fast—what we can start doing is immense.”

Willis has been in the sport for decades, much of that time as a technical director across various title-winning teams. He says collecting and understanding data is the area with the biggest gap between successful and unsuccessful F1 teams these days. So just as he helped encourage the team to embrace computer simulations and models once upon a time, today he’s also championing machine-learning adoption within Mercedes.

“I’m not sure whether to say F1 is slow to the party, but we’re just starting to apply this to a lot of areas. We have a handful of machine learning projects in very different areas: race strategy, testing, analysis of software, analysis of component failures,” he says. “Ultimately, it’ll lead to better decision-making. We have lots of data, but you have to do something to categorize it and know where it is before it becomes knowledge. When you then have that knowledge and understand how it all fits together; that’s the real driver for performance in F1.”

Listing image by Mercedes-AMG Petronas Motorsport

Read the full article here by Ars Technica

No comments: