The world’s top supercomputers may be performance powerhouses, but they’re also energy hogs, often using enough power to light up a small city.
Such power consumption prompted two professors at the Virginia Polytechnic Institute and State University to come up with a way to commend companies that build energy-efficient supercomputers. Kirk Cameron and Wu Feng, both associate professors in the Blacksburg, Va., university’s computer science department, are preparing to release the Green500 List of the most energy-efficient supercomputer installations in the world.
“The list is meant to encourage people to develop systems that use power efficiently,” said Cameron. “If you do that, you decrease costs, which are really high — between $800,000 and $1 million a year per megawatt. It’s ridiculously expensive. You’re looking at $1 million to $4 million a year, easily, to run a mega system.”
The Green500 List is a rating of the energy efficiency of machines that make the Top500 List of supercomputers, a ranking of the most powerful supercomputers in the world. The 30th installment of the biannual Top500 List is slated to be issued on Nov. 14 at the Supercomputing Conference in Reno, Nev.
The Green500 List is expected to be released just a few days after the Top500 List, giving the researchers time to rank any newcomers.
Cameron said he and Feng have put together a make-shift green list for the past few years, but it only rated some of the supercomputers ranked on the Top500 List, and then only rated them by their peak performance numbers.
This year’s list, he said, really marks its official coming out, with its attempt to rank every machine on the Top500 List. They have asked users of the Top500 machines to provide measurements based on a standard workload. Cameron explained that they hope to receive detailed information from about 100 of the supercomputers — calculating energy consumption, heat output and efficiency using the power the machine is consuming.
Machines at user sites that don’t provide the requested data will be measured by peak performance numbers. Researchers will give preference to the supercomputers that they have working data on, according to Cameron.
So, which supercomputers do the researchers find tend to be more efficient?
Cameron said the machines that use blade servers tend to be much more energy efficient than machines that cluster a bunch of servers together. He noted that the Earth Simulator, a NEC vector supercomputer being run in Japan, uses a lot of energy, producing 5.12 MFLOPS per watt. IBM’s BlueGene supercomputer, which currently sits at the head of the Top500 List, does not use blades but still produces a more efficient 187.79 MFLOPS per watt.
“I think companies will start to focus on more efficient machines because if you want to get more performance, it has to be more efficient,” said Cameron. “If not, it’s self-limiting. You can’t build a machine that needs its own power plant… Being green may just be a popular term that may fade, but energy efficiency is good business. And that should have meaning to an IT manager. Two machines can have the same performance, but if one uses half the power, then choosing that one is a business decision.”