1-1(81) 2014 MATHEMATICS AND MECHANICS
A. V. Utkin, M. S. Ozhgibesov
CUDA and MPI Programming Models for Problems of Molecular Dynamics
One of the most difficult issues in molecular dynamic modeling is known to be a large computational time required for simulation even relatively small systems of atoms. Implementation of highly efficient parallel scalable algorithms is a key for effective solution of the problem mentioned above. In the paper, three implementations of parallel MD algorithms are tested. The first approach is based on a highly parallel MPIbased program design for a computer cluster with distributed memory. The second parallel algorithm is implemented on CUDA based GPUs by NVIDIA. It should be noted that modern high performance computing systems are combinations of MPI clusters equipped with GPGPUs, which turns them into so-called heterogeneous computing clusters. In this case, MPI technology is used for internode communications, while all computations are carried out by GPUs. Thus, the third approach of parallelization discussed in this study is based on a design of CUDA-MPI algorithm. The detailed studies and comparison of all three approaches (MPI, CUDA and CUDA-MPI) are performed to define optimal parameters and conditions of each algorithm applicability.
DOI 10.14258/izvasu(2014)1.1-28
Key words: molecular dynamics, parallel computing
Full text at PDF, 278Kb. Language: Russian.
|