Microsoft Subnet An independent Microsoft community View more

Microsoft breaks petaflop barrier, loses Top 500 spot to Linux

Microsoft misses top five, aims for "bottom 500,000"

Microsoft says a Windows-based supercomputer has broken the petaflop speed barrier, but the achievement is not being recognized by the group that tracks the world's fastest supercomputers, because the same machine was able to achieve higher speeds using Linux.

It's an interesting story that demonstrates Microsoft's improving ability to run high-performance applications, while also giving Linux fans reason to smile.

Microsoft: 'We love open source'

Windows after 25 years: A visual history

The petaflop barrier was first broken in June 2008, when IBM and Los Alamos achieved a long-sought goal in the supercomputing industry, by building a Linux-based machine that could perform one thousand trillion calculations per second, a remarkable speed known as a "petaflop."

It was the fastest supercomputer in the world at the time, and remained one of just three computers to achieve petaflop speeds through June 2010, when the twice-yearly Top 500 supercomputing sites list was released.

Microsoft, meanwhile, hit its greatest heights when it briefly cracked the top 10 of the Top 500 list in November 2008 with a Windows HPC Server-based cluster at the Shanghai Supercomputer Center. The Shanghai machine achieved speeds of 180 teraflops, less than a fifth of a petaflop.

The newest Top 500 ranking came out today, and Microsoft still hasn't placed a petaflop machine on the prestigious list. But the company claims that one of the Linux-based clusters on the Top 500 list has also achieved petaflop speeds using Windows HPC Server.

The machine is Tsubame 2.0, based at the Tokyo Institute of Technology.

"We saw outstanding performance from Windows HPC Server during our Linpack benchmarking run on Tsubame 2.0," Satoshi Matsuoka, professor at the Tokyo Institute of Technology, is quoted as saying in an embargoed Microsoft press release. "It broke the Petaflop barrier and was on par with Linux at this scale. In a power-optimized configuration, it recorded over a Gigaflop/Watt, showing it is nearly three times more energy efficient than an average laptop.  We were very excited to see this level of performance given Windows applications will be an important part of our work with industry partners."

So why isn't Windows credited with breaking the petaflop barrier by the Top 500 organization? Bill Hilf, general manager of Microsoft's technical computing group, answers that question.

The Tsubame team ran their Top 500 benchmarking tests on both Linux and Windows, and the difference in performance was less than 5% but Linux did come out on top, Hilf says. Hilf attributes Linux's slim victory to the Tokyo researchers running the Linux tests on a slightly larger number of nodes. I'm not sure why the tests were run on a different number of nodes, but I will be interviewing Matsuoka at this week's SC10 supercomputing conference in New Orleans and will attempt to find out. (UPDATE: Here's Matsuoka's take on Windows and Linux)

The Tokyo Institute could submit only one test to the Top 500 group, and obviously chose the faster, Linux-based one, Hilf says. Tsubame hit speeds of nearly 1.2 petaflops, good for fourth place in the world. Only seven supercomputers broke a petaflop, so Microsoft could have entered the top ten or perhaps even the top five if Tsubame submitted the Windows run.

The fastest computer in the latest list, as expected, is a Chinese, Linux-based system that reached speeds of more than 2.5 petaflops. Tsubame, by the way, used SuSE Linux Enterprise Server 9 in previous Top 500 runs, but the latest list did not specify which flavor of Linux was used to reach a petaflop. We do know, however, that the cluster uses Intel-based HP ProLiant servers, Nvidia graphics processors and an InfiniBand interconnect.

Despite losing a top spot to Linux, Microsoft is pleased by the results, or at least appears to be publicly. Supercomputing capabilities have historically been limited to large universities, government agencies and giant research-oriented private businesses, but Microsoft believes Windows HPC Server can expand the market to a much wider range of business customers and fill Microsoft's coffers in the process.

Instead of buying what Hilf calls an "exotic" IBM Blue Gene system or a Cray supercomputer, Windows administrators who need some basic HPC functionality can plunk down $925 for a Windows HPC Server license.

Microsoft has previously claimed that Windows HPC Server is cheaper than Linux when it comes to building HPC systems. But only five of the Top 500 computers were using Windows while 459 use Linux. Microsoft officials would surely love to improve upon that number, but realize they are unlikely to ever catch Linux in this ranking.

"We're not trying to beat Linux," Hilf says. "We're not trying to be a supercomputing company. We're trying to say 'how do we mainstream all of this stuff so that HPC becomes broadly available at all levels.'"

Hilf was also pleased by the efficiency reported by Tokyo officials during Windows tests, noting the gigaflop per watt performance noted by Matsuoka.

"For us, it's not about some exotic supercomputers that are available to a small amount of users," Hilf says. "We're really interested in the bottom 500,000 computing users."

So, it appears to be a good day for Microsoft. But it would have been a better one if weren't for that pesky Linux.

Follow Jon Brodkin on Twitter.

Join the discussion
Be the first to comment on this article. Our Commenting Policies