HPC Focus

Green Supercomputing

Energy-efficient system drives climate modeling for U Maine computer science professor

• By Dian Schaffhauser
• 01/22/09

In the process of building a grid that will allow users--including members of the public--to manipulate scientific models through a Web portal, Phillip Dickens, a computer science professor from the University of Maine, discovered he could go green with the choice of supercomputer he needed for the job. In fact, to demonstrate how low the energy requirements of a supercomputer could be, he enlisted members of the university's bicycle team to power it with their pedaling.

The demonstration, captured in the video below, shows a group of stationary bicyclists on one side of the room and a monitor displaying a glacial model on the other side. "This system--with a ridiculously small amount of power--was still doing world-class calculations," said James Bailey, marketing director for SiCortex, the company that makes the brand of supercomputer Dickens purchased.

The Search for the Right Supercomputer
Dickens's search for a computer wasn't quick. It began when he received a $200,000 grant in 2007 from the National Science Foundation to purchase a supercomputer upon which to run the Scientific Grid Portal for accessing his university's computing resources. The institution already owned a couple of other supercomputers, one a Pentium-based 16-processor Beowulf cluster and the other a 208-node Mac cluster originally set up for projects funded by the military. But those systems were power-intensive, he pointed out. "They're normal supercomputers, but they require a tremendous amount of electricity and cooling in a building dedicated to supercomputing."

Dickens spent a year researching what system to buy. He wanted to get a supercomputer that wouldn't require dedicated resources to maintain and that would fit within his relatively modest hardware budget. "It takes a lot to put another supercomputer into the current facility. I didn't want to do that," he said. "I didn't want to worry about how to pay for maintenance." Plus, the institution was pushing to make the campuses as green as possible, and a traditional supercomputer purchase wouldn't necessarily satisfy that goal.

Conversations with other vendors in the space failed to impress Dickens. "I didn't feel I was getting as much for my money as I was hoping to get."

He began talking with Joe Mannino, a sales director from SiCortex, who told him their supercomputer was very low power and had a 2 millisecond interconnection time between nodes. "I thought the guy was crazy," said Dickens. "I just wrote him off."

So Mannino had clients at the Argonne National Laboratories, where Dickens had previously worked, contact him to vouch for the company. "Two people wrote me that these guys were for real and that they had one of the machines and they liked it."

Even then, it took him another six months to make the decision to buy two units from SiCortex. "In the final analysis," Dickens said, "you get a boatload more for your money with the systems I got."

Dickens purchased a model no longer available, the SiCortex SC648, which has the capacity for 648 processors, and the SC072, used as a development machine and capable of handling up to 72 processors. A major advantage of the choice involved power use: A single processor in a SiCortex computer draws just half a watt.

As Mark Blessing, SiCortex's vice president of marketing, explained, "Unlike traditional cluster systems that use PC chips and the PC architecture, our design is streamlined." Those systems use commodity chips, he said, which can do a variety of jobs but demand additional energy in the process. The computer reduces power by using fewer parts and running ultra-low power circuits on its nodes.

"Normally, if you have a 648-processor system, you have to have a nuclear power plant in the back to power and cool the guy," Dickens joked. "If I'd bought a conventional supercomputer, I would have had to keep it in special facility." This one sits in his lab. To prepare for its arrival, he said, "we swept the floor."

Rather than requiring a specialized room with heavy duty cooling, the larger system, which stands about five feet tall and three or four feet wide--"just the size of a small refrigerator," Dickens said--is cooled with an "old fan" he brought in several years ago and a 10-year-old portable air conditioner that he claims no longer works.

"It's a much more autonomous system," the professor said. "I don't have to have a boatload of system administrators messing with it. I don't have to ask them to step up the power to keep it. It was up extremely quickly--a couple of hours."

High-end Modeling with Pedal Power
Dickens wanted an original way to show the campus community just how little electrical power the new supercomputer required for operation. That's where the cycling demonstration came in. The campus biking club and local bicyclists gathered to pedal the bicycles, each of which was hooked to a little generator. "As they were peddling, they were charging a battery that was running the supercomputer," said Dickens. "They actually put too much energy into the battery and took us down for a while. They were too strong."

In the spring Dickens will teach a course on high-performance computing, which will focus on understanding current supercomputer technology, writing parallel applications, and giving students practical experience in using the systems.

A second NSF grant of $300,000 will enable Dickens, his grant collaborators, and a team of graduate and undergraduate students to develop new software to improve the transfer rate of the massive data files required for real-time modeling work. That will feed into the scientific grid project.

For the initial demonstration, Dickens ran a well known ice sheet model, based on the work of James Fastook, also in UMaine's computer science department, that shows the effect of temperature change on the ice sheets covering Greenland. Although the original model was written for a single CPU system, Dickens' team converted the code to run on multiple processors.

This type of computation is what the SiCortex system is currently best suited for--parallel processing with high communication needs. It's less suitable for applications that put more focus on fast processor speeds. But that specialization may soon change, said SiCortex's Blessing. "Our next generation is actually going to be much more broadly applicable."

With his Web portal project, Dickens said he expects to put that model and others from the university onto the supercomputer to make them available to the general public, including middle school students.

"Essentially, there's an ordinary Web server. People will get into the portal by contacting the server using the normal Internet with their standard Web browser," explained Dickens. "Once they contact the server, there will be a collection of scientific tools they can access. The server will get the requested model running on the supercomputer. As the supercomputer is running, it will generate the data that we want to visualize. The supercomputer will send data to the visualization system, which is just a laptop with a reasonably powerful graphics card. A rendering engine will generate the images, send them back to server, and then send them back to the client who's logged on."

Users will be able to change the parameters--such as temperature or carbon emissions--and get immediate feedback via the images about whatever is being modeled.

"I'm really excited about the educational outreach component of these grants," he said. "Part of educational outreach is to disseminate important research results to the wider community, and the scientific grid portal provides an excellent platform with which to do so."