Parallel Computation Research About Molecular Dynamics Based On CPU-GPU Cluster

Molecular dynamics simulation is a kind of discrete simulation method; it is also a simulation method for molecular motion which needs a large amount of computation. The method is mainly based on the laws of molecular movement, calculating the change of molecular movement within a certain period of time, and we could obtain a variety of physical properties and chemical properties of molecular systems through the qualitative and quantitative analysis to the data from the calculation. It has been widely used in physics, chemistry, biology, materials, medicine and other fields.GPU is designed for intensive and highly parallel computation. The application of CUDA reduces the development effort of GPU's general computation. GPU's computing power is quite powerful; however to the meaningful ultra-large-scale simulation system the molecular weight can attain to tens of millions or even to hundreds of millions, for these systems, the required computing power is far greater than a single GPU, so we can see that there is research and practical significance to develop multi-GPU parallel computing architecture.This paper is mainly about parallel computing research for the large-scale molecular dynamics simulation system. Build a parallel computing platform-CPU-GPU cluster; In the cluster, GPU is worked as a coprocessor, while parallel numerical computations should be accomplished by GPU, the remaining operations is the CPU's duty. We accomplish molecular dynamics simulation on this cluster using the CUDA development environment. We propose a domain decomposition method to divide the parallel tasks named "minimum surface area", which is suitability for GPU, and this method is a static load balancing algorithm and reduces the communication cost of inter-nodes. The cell-list method is used to search for adjacent particles and it is conducive for parallel computation on the GPU Using MPI message communication mechanism, and proposed a suitable communication method for the designed system. Comprehensive considered the algorithms above we implement the parallel design which is suitable for CPU-GPU cluster.We can safely make the conclusion that GPU's computing power is much greater than CPU and the CPU-GPU cluster's computational efficiency is much larger than that of CPU cluster for large-scale the molecular system's computation. We also design and implement molecular parallel system based on FPGA, GPU also has some advantages compared with the FPGA implementation from the experiment results. For large-scale molecular systems which have millions, or tens of millions particles, the CPU-GPU cluster's computational efficiency is much larger than that of CPU cluster.