SCIENCE & ENGINEERING NEWS
San Diego, CALIF. — Digital magnetic recording is the central technology of information storage. Magnetic recording systems include the storage media (disks and tapes) and the transducers that write and read information: the record and playback heads. Disks, for example, are layered structures of thin films of cobalt-based crystals, in which the density of the stored information depends largely on the size of the grains in the coating. The smaller the grains, the more bits can be recorded in a given area. As grain sizes have become smaller, the capacity of hard disk drives has doubled annually for the past few years. But when bits are written on tinier and tinier grains, the stored information is eventually subject to thermal instability – a tiny kick in energy can change a “1” to a “0.”
UC San Diego scientist H. Neal Bertram and colleagues, in collaboration with SDSC staff, are exploring models of computer disks to determine the thermal stability limits for ultrahigh-density systems with extremely fine grains. Their work is stretching the capacity of the Sun HPC 10000 at SDSC.
“Our group focuses on the fundamental physics of magnetic recording,” said Bertram, professor of electrical and computer engineering, who holds an Endowed Chair in the Center for Magnetic Recording Research (CMRR) at UC San Diego. “We use sophisticated, computationally intensive models to simulate the recording process in advanced devices, taking into account a variety of variables, from short- and long-range interactions among grains to inhomogeneities in grain microstructure that affect recording results.”
The CMRR was founded on the UC San Diego campus in 1983 by a consortium of 12 U.S. companies in the magnetic recording industry. Bertram and his students have used SDSC resources since 1986. “Now we are on the threshold of a great increase in the size and power of our calculations,” Bertram said, “thanks to massively parallel computation and our collaboration with the computational scientists at SDSC.”
The group has used supercomputing resources to analyze the time evolution of recording on disks. Most current disk technologies employ longitudinal recording, with the magnetization direction of each grain in the plane of the medium. However, perpendicular recording, with the magnetization aligned vertically to the disk surface, may be coming into use and is currently under serious investigation by all the major disk drive companies. Perpendicular recording appears capable of main taining good signal-to-noise ratio and thermal stability at higher densities. Another candidate is patterned media, in which each bit is assigned to an individual grain. “We have used our codes to analyze all types of media,” said Bertram, who has been an IEEE Distinguished Lecturer and is the author of Theory of Magnetic Recording (Cambridge University Press, 1994).
Current disk technology permits areal densities of tens of gigabits per square inch, which can be recorded or read out at rates as high as 200 MHz. The industry’s next objective is to go an order of magnitude higher in density and speed, with hundreds of gigabits per square inch readable at gigahertz rates. Ideally, new media should exhibit the same 10-year thermal stability at room temperature as current media.
In modeling the stability of magnetic disk media, Neal Bertram and Hong Zhou simulated the difference between two states of affairs. Above is a representation of the medium at the beginning of a run on the Sun HPC 10000 at SDSC. Red and blue colors represent two magnetization directions. The thermal decay was exaggerated slightly to get this contrast. At the start of the simulation, bit striping is clearly evident.
Recently, as part of an NPACI Strategic Applications Collaboration (SAC) project, Amitava Majumdar of the SDSC Scientific Computing Department worked with Hong Zhou, Bertram’s graduate student, on a simulation of thermal instability in dense disk media, using smaller and smaller grain sizes.
“As the volume of each grain decreases, the amount of energy needed to reverse the direction of magnetization also becomes smaller,” Zhou said. “We use a Monte Carlo method to randomly distribute and change magnetization directions for an assembly of grains and calculate the probability that such changes will occur within a given time – from a nanosecond to 10 years.”
Typically, the medium is represented as a 128 x 128 array of hexagons. A single run typically involves 120 calculations with varying random properties (grain orientations) to determine medium noise and thus the system’s signal-to-noise ratio. When the collaboration began, Zhou was obtaining a speed of about 250 Mflops on a single processor of the Cray T90 at SDSC.
“But a single run on one such processor took about 18 hours of processor time,” Majumdar said. “Our analysis indicated that the code might perform very efficiently on the Sun HPC 10000.” The HPC 10000 is a parallel machine with 64 processors and 64 GB of shared memory uniformly accessible from each processor. Majumdar and Zhou developed an implementation of the code using the Message-Passing Interface (MPI) library that was parallel across the initial conditions. “With this change and a number of smaller improvements in the code, the run time on 40 processors of the Sun was reduced to 2.5 hours. The parallel code scales almost linearly, since initial conditions can be simulated independently of one another.”
“That speedup alone is very important for us,” Zhou said. “Professor Bertram and I can now test more initial conditions, since the code is parallel, and we can increase the size of our simulation, because of the large memory of the HPC 10000, to represent higher areal bit densities.” Majumdar and Zhou have further improved the code by finding and linking mathematical libraries for parallel random number generation. “We are also investigating changes in the Monte Carlo code that may reduce the variance and thus help to simulate more realistic problems faster,” Majumdar said.
SAC program coordinator Bob Sinkovits and SDSC scientist Stuart Johnson have begun work with another Bertram student, Chris Seberino, on the group’s code for analyzing tape recording. “The main part of this code is an implementation of the Fast Multipole Method,” Johnson said. “We are adapting an efficient parallel implementation of this method that was already in use by scientists in another area of physics entirely.”
Sinkovits pointed out that the ability of the SAC team to improve the efficiencies of major codes benefits from the synergies implicit in working across multiple disciplines. “We become code mavens – people with a kind of expertise rarely developed in academic groups within a single discipline,” he said. “The more experience our group has with this kind of work, the better our unique form of collaboration becomes, for all kinds of science.” For example, NPACI has a range of high-end machines with architectures and protocols that differ. One code with certain demands on machine memory may run better on the Sun HPC 10000 than the IBM SP machines; another, with differing characteristics, may run better on one of the SP configurations.
“I am very much impressed with what the SDSC computational scientists have been able to do with our codes,” Bertram said. “We’ve been given some new handles on our ever-changing problems that can increase our power to predict and guide the way forward for a vital technology.”