November 19, 2008
A team led by Thomas Schulthess of Oak Ridge National Laboratory (ORNL) has broken the petaflop barrier with a supercomputing application likely to accelerate the revolution in magnetic storage.
Using ORNL's upgraded Cray XT Jaguar supercomputer, the team was able to achieve a sustained performance of 1.05 quadrillion calculations a second, or 1.05 petaflops, for an application that simulates the behavior of electron systems. Jaguar itself was recently upgraded to a peak performance of 1.64 petaflops, making it the world's first petaflop system dedicated to open scientific research. The team's simulation ran on nearly 150,000 of Jaguar's 180,000-plus processing cores.
Among its benefits, the application promises to advance scientific understanding of magnetic devices such as computer hard drives. In the last couple of decades, hard drive storage capacity has grown at an extraordinary rate. The associated risk, though, is that with increasing storage density, these amazing devices tend to become less stable.
Hard drives hold information by magnetizing tiny regions of a platter, with regions magnetized in one direction counting as ones and in the opposite direction as zeroes. With the exponential growth of storage capacity, these miniscule spots have gotten progressively even smaller; and the smaller the spot, the more likely its magnetic direction is to be incorrectly and unexpectedly reversed. Since disorder at the atomic scale increases with temperature, a hard drive kept as warm as room temperature becomes increasingly susceptible to random changes -- meaning lost data -- as storage density rises.
"A big problem in magnetic recording is that as you make the bits smaller and smaller, thermal excitation will essentially randomize them and you will lose information," explained Markus Eisenbach of ORNL. "If that happens in 500 years you don't care, but if it happens tomorrow you're really unhappy."
The team's current approach differs fundamentally from earlier efforts because it is able to set aside empirical models and their attendant approximations to tackle the system through first-principles calculations. Eisenbach, who serves as the team's developer for the project, noted that this empirical approach was far too computationally intensive for earlier computer systems.
"It's the new Jaguar coming on line that makes it really feasible," he said. "If you have a classical Heisenberg model, an energy calculation takes perhaps milliseconds. For this first-principles calculation, an energy calculation takes tens of seconds. So it's orders of magnitude slower. You really need a computer of that size."
The team simulates the effect of heat on a magnetic material by combining two methods. The first -- known as locally self-consistent multiple scattering, or LSMS -- describes the journeys of scattered electrons by applying density functional theory to solve the Dirac equation, a relativistic wave equation for electron behavior. The code has a robust history, having been the first code to run at a sustained trillion calculations per second and earned its developers the prestigious Gordon Bell Prize in 1998.
The shortcoming of this approach, though, is that it is used primarily to describe a system in its ground state at a temperature of absolute zero, or nearly 460°F. In order to include the energy brought to the system by temperatures outside a laboratory freezer, the team's simulations incorporate a Monte Carlo method known as Wang-Landau, which guides the LSMS application to explore electron behavior at a variety of temperatures.
According to Eisenbach, the two methods are ideally suited to massively parallel computing systems. They scale linearly, meaning the need for computing resources grows at the same rate as the size of the system being simulated, and LSMS can be scaled to very large materials systems by assigning one atom to each processing core.
As a result, the team is able to use the petascale Jaguar system to simulate nanoparticles approaching technologically interesting sizes.
"We're really getting to a size where you could do calculations for nanoparticles that are also the focus of experiment," Eisenbach noted. "Experiments come from large systems and manage to get smaller and smaller, and we are coming from just a few atoms and getting to the point where experimentally accessible sizes and computationally accessible sizes meet."
He would not predict what the project will find, since the team is taking a new approach to the problem. Nevertheless, he noted that hard drive manufacturers are watching this issue closely; as hard drive-capacity continues to grow, the importance of a more complete understanding of magnetic materials will also grow.
"The idea is to find materials that make it sufficiently hard for random temperature fluctuations to turn the bits around, so the information is still on your hard disk when you look at it next year. We have been talking with people at hard disk manufacturers. Certainly, it's an important issue that gets discussed at magnetism conferences."
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?