Elsevier

Computer Physics Communications

Volume 211, February 2017, Pages 73-78
Computer Physics Communications

Crystal MD: The massively parallel molecular dynamics software for metal with BCC structure

https://doi.org/10.1016/j.cpc.2016.07.011Get rights and content

Abstract

Material irradiation effect is one of the most important keys to use nuclear power. However, the lack of high-throughput irradiation facility and knowledge of evolution process, lead to little understanding of the addressed issues. With the help of high-performance computing, we could make a further understanding of micro-level-material. In this paper, a new data structure is proposed for the massively parallel simulation of the evolution of metal materials under irradiation environment. Based on the proposed data structure, we developed the new molecular dynamics software named Crystal MD. The simulation with Crystal MD achieved over 90% parallel efficiency in test cases, and it takes more than 25% less memory on multi-core clusters than LAMMPS and IMD, which are two popular molecular dynamics simulation software. Using Crystal MD, a two trillion particles simulation has been performed on Tianhe-2 cluster.

Introduction

Nuclear power is getting more and more concerned as a clean energy resource. To use the nuclear power properly, material irradiation effect in nuclear reaction is one of the most important keys. Materials irradiation effect mainly refers to the interaction of neutrons, charged particles or electromagnetic rays and other radiation with solid materials product. Study on materials irradiation effect involves many research fields, such as nuclear reactor structural material, ion probe, plasma processing and ion modification.

The material problem is a major factor restricting the development of nuclear power. The material radiation damage crosses 9 orders of magnitude, from the nanometer’s atomic scale to the meter’s macroscopic scale; as well as from bond breaking processes in the time scale of picosecond to nonlinear process of engineering structural failure and destruction in the time scale of decades. Because of the lack of the high-throughput irradiation facility and difficulty in observing the process of evolution, we cannot get well understanding of the internal microstructure evolution process and the mechanism of radiation effect on materials. In other words, a basic physical understanding about atomic scale materials design and development guidance under irradiation environments lacks  [1].

The development of the high-performance computer technology can make such a simulation on materials irradiation effect possible, thus, it can provide an understanding from micro-level to reveal the microstructure evolution of material failure, and develop a quantitative relationship between macro properties and microscopic processes  [2], [3], [4], [5].

Computational material science has become one of the most important research fields of materials science because the computer simulation can offer us an insight into the details of materials and reduce the costs. The combination of radiation experiments and computer simulations cannot only improve research efficiency, but also greatly reduce the human and financial resources cost  [6], [7]. Molecular Dynamics (MD) is a thermodynamic calculation method based on the theory of Newtonian mechanics; it can be used in the ensemble calculations such as NPT, NVE, NVT. Based on solving the equations of motion of all the particles in the system, thermodynamic quantities and other macroscopic properties of system can be obtained. MD simulation has been widely used in various fields of physical, chemical, biological, materials, medicine, etc.

One can note that the storage capacity constraints the scale of MD simulation from tens of thousands to tens of millions of atoms. However, the typical material microstructure and the defect size are much larger than such a calculated scale. Thus, to expand the scale of MD simulation is significant in material science. To such an aim, we can develop a new method with using less memory or increase the storage capacity on multi-core clusters. Both can increase the scale of MD simulation.

Recently, with the rapid development of the high performance computer technology, MD simulation can be performed in larger scale simulation. New MD software based on hybrid architecture high performance computer has been developed. A piece of MD simulation software facing high performance computer—ddcMD has been firstly developed by LLNL and IBM  [8]. A piece of massive parallel MD simulation software—LAMMPS has been developed by Sandia national laboratory  [9]. University of Stuttgart developed a piece of MD simulation software in 1997, named IMD  [10], [11], [12], and in 1999, it achieved the world biggest record with 5109 atoms. Based on IMD, University of Stuttgart developed a new piece of MD simulation software—ls1 Mardyn  [13]. It holds the world biggest MD simulation record with 41012 molecules.

In this paper, a new data structure has been designed for parallel MD simulation. It focuses on the crystal structure characteristics of the body-centered cubic (BCC) metallic materials without the neighbor list or the linked cell. Based on this data structure, we developed the molecular dynamics simulation software named Crystal MD  [14]. To take advantage of characteristics of the BCC structure, array elements’ location in Crystal MD can be used directly to reflect the atoms’ spatial arrangement. Compared with the traditional MD data structure, this new data structure can reduce memory usage per atom, which results in larger scale MD simulations on clusters with same memory capacity. Compared with the current mainstream in large-scale MD simulation open source software IMD and LAMMPS, the Crystal MD can significantly expand the scale of simulation for the crystal with BCC metal materials.

The paper is organized as follows. Section  2 generally introduces Crystal MD. Section  3 proposes the new MD simulation data structure. Section  4 presents a new communication method based on the new data structure. Performance analysis and discussion are given in Section  5. Conclusions and future work are drawn in Section  6.

Section snippets

Crystal structure

Mechanical properties of materials are dedicated by its’ microstructure  [15]. Structural materials in reactor systems are predominantly crystalline, metallic alloys. Neutron radiation can displace atoms from their lattice sites and the formation of point defects, and the migration and clustering of point defects can cost great damage on its mechanical properties  [16]. In Crystal MD, as we are aiming at studying the radiation effect of alloy metal with the BCC structure, we focus on MD

Data structure in Crystal MD

The design of the data structure in MD simulation determines the efficiency of finding neighbors during the force calculation. Furthermore, it determines the scale that the computer simulation can reach. Two data structures are commonly used in MD simulation software: Linked cell  [21], [22], [23], and Neighbor list  [24]. Linked cell data structure divides the simulation volume into equally sized cubic cells, which have an edge length equal to the cut-off radius. Therefore, this data structure

Communication scheme in Crystal MD

For developing massively parallel MD simulation with Crystal MD, we use standard domain decomposition to divide the simulation box, which means each process gets same volume of box. The scenario that the metals with BCC structure can be simulated by Crystal MD, it is because the feature of BCC structure determines its spatial distribution is almost fixed during the whole simulation. This requires that the atoms at lattice points, which are needed to be sent to neighbor processors as their ghost

Performance analysis and discussion

The experiments are performed on large-scale parallel computers to test the performance of Crystal MD, which focus on scalability and memory usage. In order to prove the superiority of the proposed data structure in the MD simulations, we compared the performance of simulations with LAMMPS (version 20150320) which uses neighbor list data structure and IMD (version 20140121) which uses Linked cell data structure.

We have deployed Crystal MD on both supercomputing clusters: Era and Tianhe-2.

“Era”

Conclusions

Based on the analysis of the calculation for finding neighbor atoms during the MD simulation, we present the new MD software—Crystal MD using a new proposed data structure named as lattice neighbor list. We compared the memory usage with other MD software to analyze the constraint on MD simulation scale. For metals with the BCC structure, Crystal MD saves much more memory to improve the scalability of MD simulation at the same memory capacity comparing to LAMMPS and IMD. We have performed a two

Acknowledgments

The research is partially supported by the Hi-Tech Research and Development Program (863) of China No. 2015AA01A303, Natural Science Foundation of China under Grant No. 61303050 and the Youth Innovation Promotion Association, CAS(2015375).

References (26)

  • C. Hu et al.

    Future Gener. Comput. Syst.

    (2016)
  • Jue Wang et al.

    Future Gener. Comput. Syst.

    (2016)
  • Jue Wang et al.

    Future Gener. Comput. Syst.

    (2010)
  • Li Shigang, Zhang Yuanquan, Xiang Chunyang, Shi Lei, Fast convolution operations on many-core architectures, in: High...
  • Li Shigang et al.

    Cluster Comput.

    (2014)
  • Jilin Zhang et al.

    Future Gener. Comput. Syst.

    (2015)
  • Zhang Jilin et al.

    Appl. Math. Inf. Sci.

    (2013)
  • F.H. Streitz, J.N. Glosli, M.V. Patel, et al. 100+ TFlop Solidification Simulations on BlueGene/L, in: SC05,...
  • LAMMPS manual,...
  • J. Stadler_ et al.

    Internat. J. Modern Phys. C

    (1997)
  • J. Roth

    Proceedings of the Workshop Md on Parallel

    Imd–A Molecular Dynamics Program and Applications

    (1999)
  • Large systems 1: molecular dynamics; http://www.ls1-mardyn.de/  (accessed...
  • Cited by (0)

    View full text