Does being first in supercomputing still matter?
Source: Patrick Thibodeau
Supermoon as seen from Jamnagar, Gujarat, India on June 23, 2013 Credit: Image credit: Flickr/SONARA ARNAV
The goal of building an exascale system is compared to the U.S. effort to go to the moon
The European forecast of Hurricane Sandy in 2012 was so far ahead of U.S. models in predicting the storm's path that the National Oceanic and Atmospheric Administration was called before Congress to explain how it happened.
NOAA administrator Kathryn Sullivan told lawmakers at a hearing last year that the Europeans "set a target and a policy of staying very close to the leading edge of computational capacity." That's in contrast to the U.S., which "falls further and further behind the cutting edge" and then follows it with a "big step forward," she said.
That's how things work in the U.S. When the Soviets launched Sputnik in 1957, the "big step forward" for the U.S. was the moon landing in 1969. It crushed the competition.
Until recently, the U.S. has never faced global competition in high performance computing. Its leadership was assured in hardware and software development. But America's leadership, while still intact, is being challenged.
Supercomputers have become high priority investments in Europe, Japan and China, which operates the world's fastest system. The U.S. has set a goal of 2023 as its delivery date for an exascale system, and it may be taking a risky path with that amount of lead time.
President John F. Kennedy delivered his famous "we choose to go to the moon" speech in 1962, and seven years later a man walked on the moon. The U.S. exascale goal is nine years away.
One thing that has given U.S. policymakers comfort is that many of the components that make up high-performance computing systems are American made. But this vendor landscape is about to see a major shakeup.
IBM's decision to sell its x86 server business to Lenovo will turn the China-based company, in short order, into one of the largest HPC vendors in the world, according to IDC.
"Lenovo may become the number two HPC provider literally by the end of this year," said Earl Joseph, an analyst at IDC. Hewlett-Packard is number one. If not in the second position, Lenovo will be close to it.
When Sputnik circled the Earth, it was clear that the Russians had made a significant accomplishment. But if the Europeans or China beat the U.S. to an exascale system, a 1,000 petaflop machine, it won't be readily clear that they have achieved true exascale performance.
The view here at the SC14, the major supercomputing conference, is that petaflops, as a benchmark measured by the performance of the Linpack calculation, will not matter with an exascale system.
If the first exascale machine arrives offshore, there's a good chance it will immediately be labeled in the U.S. as a "stunt machine," capable of setting a new Linpack benchmark, but not very useful for running science applications. The U.S. goal is to build an exascale system that can not only run science applications, but can also deliver marketable technologies to boost the U.S. tech industry.
Countries will still want to brag that they achieved exascale first, but unlike the Cold War that helped to give rise to Sputnik, there is "strong collaboration" with Japan and the European Union on exascale software development, and efforts with China are improving, said William Harrod, research division director for the U.S. Department of Energy's Advanced Scientific Computing Research program, at a conference forum Wednesday.
The strong collaboration on software doesn't replace national competitive strategies, or the goals of Europe and China, in particular, to use the push to exascale to improve their own tech industries.
The U.S. appears set to move at its own pace, aware of what's going on globally but blind to it at the same time, but this could be dangerous. Take the NOAA.
Cliff Mass, a professor of meteorology at the University of Washington, wrote on his blog last month that the U.S. is "rapidly falling behind leading weather prediction centers around the world" because it has yet to catch up in computational capability to Europe.
That criticism followed the $128 million purchase of a 16-petaflop Cray supercomputer by the U.K.'s Met Office, its meteorological agency.
NOAA recently put into production two 213 teraflop systems.
Was the moon landing just a stunt? Some say it was, according to Rick Stevens, associate laboratory director at the Argonne National Laboratory in Illinois, at a forum here on exascale. Others argued that it was our "greatest achievement."
Whatever the moon landing was, it made a point.
| }
|