Portal de Periódicos da CAPES

Accelerating Molecular Dynamics Simulations on PlayStation 3 Platform Using Virtual-GRAPE Programming Model

2008; Society for Industrial and Applied Mathematics; Volume: 30; Issue: 6 Linguagem: Inglês

10.1137/070692054

1095-7197

Tetsu Narumi, Shun Kameoka, Makoto Taiji, Kenji Yasuoka,

Distributed and Parallel Computing Systems

Molecular dynamics (MD) simulation requires huge computational power because each atom interacts with another by long-range forces, such as Coulomb and van der Waals forces. Therefore, parallel computers or special accelerators, such as MDGRAPE-3, are required for accelerating MD simulations. A video game processor in a Sony PlayStation 3 or NVIDIA's graphic accelerator card is also a candidate hardware for accelerating MD simulations, because the peak performance of the latest video game processors exceeds that of a current PC's CPU, and they are also very cost-effective. However, the software development for these processors requires much more time compared to CPUs of PCs because the hardware has a lot of parallel processing. We propose the virtual-GRAPE programming model to utilize the hardware resource of video game processors with minimum time for software development. GRAPE is a special-purpose computer used to accelerate particle-based simulations: astrophysical or MD simulations. Under the Virtual-GRAPE model, the subroutine whose calculation speed is accelerated by the special hardware, GRAPE, is replaced with a specially tuned subroutine to be used without the accelerator. We implemented this model in a PlayStation 3 to accelerate the “sander” MD module in the AMBER software package. We were able to achieve an acceleration of 20 times, compared to a serial job using an Intel Xeon 5160 processor. Its performance cost is far superior to that of a PC or an MDGRAPE-3. To obtain the highest performance from the subroutine, most of the arithmetic operations in the tuned routine were performed with single precision accuracy, which is sufficient for MD simulations.