A team led by Thomas Schulthess of the U.S.
Department of Energy's Oak Ridge National Laboratory received the prestigious
2008 Association for Computing Machinery (ACM) Gordon Bell Prize Thursday after
attaining the fastest performance ever in a scientific supercomputing application.
Schulthess is group leader of ORNL's Computational Materials Science Group
and recently accepted a position as director of the Swiss National Supercomputing
Center at Manno, an institution of ETH Zurich. He and colleagues Thomas Maier,
Michael Summers and Gonzalo Alvarez, all of ORNL, achieved 1.352 quadrillion
calculations a second--or 1.352 petaflops--on ORNL's Cray XT Jaguar supercomputer
with a simulation of superconductors, or materials that conduct electricity
without resistance. By modifying the algorithms and software design of its DCA++
code to maximize speed without sacrificing accuracy, the team was able to boost
performance tenfold with the help of John Levesque and Jeff Larkin of Cray Inc.
Jaguar was recently upgraded to a peak performance of 1.64 petaflops, making
it the world's first petaflop system dedicated to open research. The team's
simulation made efficient use of 150,000 of Jaguar's 180,000-plus processing
cores to explore electrical conductance.
To put the achievement into perspective, it would take every man, woman and
child on earth more than 500 years to work through as many calculations as DCA++
gets through in a single day--and that's assuming each of us worked day and
night solving one calculation a second.
Researchers have known about superconductors for nearly a century and have
prized these materials both for their ability to conduct electricity without
resistance, or energy loss, and for their especially strong magnetic field.
Superconducting materials have obvious potential application in power transmission,
and superconducting magnets have found a place in hospital magnetic resonance
imaging machines, particle accelerators such as Europe's Large Hadron Collider,
and magnetic levitation transportation systems.
The challenge is that superconducting materials must be very, very cold. Even
so-called high-temperature superconductors--discovered in the mid-1980s--must
be chilled to a "transition temperature" of around ?°F before
they exhibit their amazing behavior. In addition, a full scientific explanation
is missing of how high-temperature superconductors work.
The team used the DCA++ application within a promising mathematical framework
known as the two-dimensional Hubbard model. These simulations were the first
in which it had enough computing power to move beyond ideal, perfectly ordered
materials. By looking at materials with disorder--or impurities--the team is
moving toward the necessarily imperfect materials found in the real world.
"The real materials are very inhomogeneous," noted team member Thomas
Maier of ORNL.
Specifically, the team focused on chemical disorder in high-temperature superconductors
known as cuprates--layers of copper oxide separated by layers of an insulating
material. By advancing our understanding of the interplay between these imperfections
and superconductivity, the work promises to help researchers push transition
temperatures ever higher, possibly approaching the lofty goal of "room-temperature
superconductors," or materials that exhibit this behavior without artificial
The team studied the local repulsion between electrons on the same atom. Because
electrons have a negative electrical charge, they push one another away in what
is known as a Coulomb repulsion. For the material to become superconducting,
however, the electrons must overcome this repulsion and join into units called
Cooper pairs. The team is looking to take advantage of an earlier discovery
that indicates the insulating material promotes this process by drawing electrons
away from the copper oxide layer.
"If you draw electrons away from the copper oxide layers, they become
superconducting," Maier said. "Then the question is, what happens
if you replace lanthanum with strontium, for instance. You do have different
potentials, but you should also have different Coulomb repulsions on each site."
To achieve the sustained speed demonstrated in the simulation, the team made
two fundamental changes to the DCA++ application, allowing it to delay memory-intensive
operations and use a less memory-intensive data form. Both of these techniques
exploit the fact that DCA++ uses the Monte Carlo approach, which relies on random
sampling of a variable to explore systems such as the two-dimensional Hubbard
model that do not lend themselves to an exact solution.
Between the two approaches, the team was able to boost the speed of the application
by a factor of about 10, according to team member Marcus Eisenbach of ORNL's
National Center for Computational Sciences. This increase in speed allows the
team to look at a wider variety of materials in increased detail.