high performance computing on graphics processing units: hgpu.org

The Poisson-Boltzmann equation models the electrostatic potential generated by fixed charges on a polarizable solute immersed in an ionic solution. This approach is often used in computational Structural Biology to estimate the electrostatic energetic component of the assembly of molecular biological systems. In the last decades the amount of structural data concerning proteins and other biological macromolecules has remarkably increased, as it can be seen by looking at the Protein Data Bank. In order to fruitfully exploit these data a huge computational power is needed as well as the availability of software tools capable to exploit it. It is therefore necessary to move towards High Performance Computing, and to develop proper parallel implementations of already existing and of novel algorithms. Nowadays, even workstations can provide an amazing raw computational power: up to 10 TFLOPS on a single machine equipped with multiple CPUs and accelerators such as the Intel Xeon Phi or GPU devices. The actual obstacle to the full exploitation of modern heterogeneous high performance computing resources is therefore efficient parallel coding and porting of software on such architectures. In this paper, we propose the implementation of a full Poisson-Boltzmann solver based on a Finite-Difference scheme using different and combined parallel schemes, and in particular a mixed MPI-CUDA implementation. This approach includes MPI-CUDA as well as multicore CPUs, multi GPUs and clusters of multi GPUs. Results show great speedups when using the two schemes, achieving a 18.9 speed up using three GPU cards compared to the serial version of the code.