How Cities Use HPC at the Edge to Get Smarter 

Cities are sensoring up, collecting vast troves of data that they run through predictive models to address problems that, in some cases, city managers didn’t even know existed.

Speaking at SC17 in Denver this week, a panel of smart city practitioners shared the strategies, techniques and technologies they use to understand their cities better and to improve the lives of their residents. With data coming in from all over the urban landscape and worked over by machine learning algorithms, Debra Lam, managing director for smart cities & inclusive innovation at Georgia Tech who works on strategies for Atlanta and the surrounding area, said “we’ve embedded research and development into city operations, we’ve formed a match making exercise between the needs of the city coupled with the most advanced research techniques.”

Panel moderator Charlie Cattlett, director, urban center for computation & data, Argonne National Laboratory, works on smart city strategies for Chicago, said that the scale of data involved in complex, long-term modeling will require nothing less than the most powerful supercomputers, including the next generation of exascale systems under development within the Department of Energy. The vision for exascale, he said, is to build “a framework for different computation models to be coupled together in multiple scales to look at long-range forecasting for cities.”

“Let’s say the city is thinking about taking 100 acres and spend a few hundred million dollars to build some new things and rezone and maybe augment public transit,” Cattlett said, “how do you know that that plan is actually what you think it’s going to do? You won’t until 10-20 years later. But if you forecast using computation models you can at least eliminate some of the approaches that would be strictly bad.”

With both Amazon and Microsoft in its metropolitan area, it’s not surprising that Seattle is doing impressive smart city work. Michael Mattmiller, CTO of Seattle, said good planning is necessary for a city expected to grow by 32 percent. Mattmiller said 75 percent of the new residents moving to Seattle are coming for jobs in the technology sector, and they will tend to have high expectations for how their city uses technology.

Some of Seattle’s smart city tactics are relatively straightforward, if invaluable, methods for city government to open the lines of communication with residents and to respond to problems faster. For example, the city developed an app called “Find It, Fix It” in which residents who encounter broken or malfunctioning city equipment (broken street light, potholes, etc.) are encouraged to take a cell phone picture and send a message to the city with a description of the problem and its location.

Of a more strategic nature is Seattle’s goal of becoming carbon neutral by 2050. The key challenges are brought on by the 100,000 people who come to the downtown areas each day for their jobs. The city’s Office of Sustainability collects data on energy consumption from sensors placed on HVAC and lighting systems in office buildings and retail outlets and has developed benchmarks for comparing energy consumption on a per-building basis, notifying building owners if they are above or below their peer group.

Mattmiller said Amazon and Microsoft helped build analytics algorithms that run on Microsoft Azure public cloud. The program is delivering results; Mattmiller said energy consumption is down, with a reduction of 27 million tons of carbon.

Seattle also analyzed weather data and rainfall amounts, discovering that the city has distinct microclimates, with some sections of the city getting as much as eight more inches of rain (the total annual amount of rain in Phoenix) per year than others. This has led to the city issuing weather alerts to areas more likely to have rain events and to send repair and maintenance trucks to higher risk areas.

Transportation, of course, is a major source of pollution, carbon and frustration (30 percent of urban driving is spent looking for parking spaces). Seattle trolled resident for ideas and held a hackathon that produced 14 prototype solutions, including a team from Microsoft who bike to work: they developed a machine learning program that predicts the availability of space on bike racks attached to city buses, “an incredibly clever solution,” Mattmiller said.

In Chicago, Pete Beckman, co-director, Northwestern Argonne Institute of Science and Engineering, Argonne National Laboratory, helped develop sensors placed throughout the city in its Array of Things project. He said that while most sensors used by cities are big, expensive and sparse, Beckman said the project managers wanted to “blanket the city with sensors,” which would collect a broad variety of data and also have significant computational power – a “programmable sensor” that doesn’t just report data but one for which you can write programs to run in the device. They also wanted it to be attractive, so students at the Art Institute of Chicago were recruited to help design the enclosure.

“This becomes a high performance computing problem,” Beckman said. “Why do you need to run programs at the edge? Why run parallel computing out there? Because the amount of data we want to analyze would swamp any network. The ability to have 4K cameras, to have hyperspectral imaging, to have audio, all that (data) can’t be sent back to the data center for processing, it has to be processed right there in a small, parallel supercomputer. Whether it’s Open CV (Open Source Computer Vision Library), Caffe or other deep learning framework like Tensorflow, we have to run the computation out at the edge.”

One scenario outlined was of a sensor detecting an out-of-control vehicle approaching a busy intersection; the sensor picks up on the impending danger and delays the pedestrian “WALK” sign and turns all the traffic lights in the intersection red. These are calculations that require HPC-class computing at the street corner.

Chicago is using its Array of Things sensors in other critical roles, such as real time flood monitoring, for tracking pedestrian, bicycle, car and truck traffic and predictively model accidents

“The questions for us in the parallel computing world,” Beckman said, “are how do we take that structure on our supercomputers and scale it in a way so we have a virtuous loop to do training of large-scale data on the supercomputer and create models that are inference-based, that are quick and fast, that can be pushed out to parallel hardware accelerated out on the edge? The Array of Things project is working on that now.”