| High-performance computing provides performance support for new applications in the field of material simulation.Combining with artificial intelligence technology,it frees material simulation from the dilemma of not being able to combine precision and efficiency at the same time.However,material simulation applications have many environmental dependencies,the compilation and installation process are relatively complex,especially after the introduction of deep learning framework into the field of material simulation,its basic environment becomes more difficult to maintain,and even between different versions of the same application,there are problems such as data incompatibility,and the experimental results are often difficult to reproduce.Container technology can solve many problems such as chaotic environment and conflicting software dependencies,but there are few cases of parallel computation among multiple containers for materials simulation applications,and there is also a lack of relevant research on computational efficiency.In addition,the traditional job scheduler is not able to achieve multi-container scheduling,and the container scheduling service cannot perform batch container scheduling in the form of jobs,and the utilization of multiple heterogeneous computing resources is not sufficient.To address the above problems,this thesis investigates a Kubernetes container orchestration service based on Docker containers,and builds a material simulation container batch processing system by combining Volcano batch container scheduler on this basis.This system can batch schedule containers in the form of jobs,so as to complete the material simulation process in a multi-container environment.First,this thesis conducts benchmark performance analysis in two containerized environments,Docker and Singularity.Perform performance tests on the CPU and network I/O of the two mainstream container scheduling services Docker Swarm and Kubernetes in a multi-container environment,and based on the actual simulation process of the typical material simulation application VASP,sort out the container orchestration service in the current batch of containers Problems and defects in the scheduling scenario;Secondly,to address the above problems and defects,we propose a solution of combining Volcano batch container scheduler on Kubernetes to realize a container batch system based on classical potential function material simulation.According to the molecular power of classic potential functions such as AIREBO and Reax FF on carbon materials in the system Learn the simulation results,analyze the feasibility and advantages of the system for material simulation;Finally,a deep learning environment is integrated into the above system to implement a container batch processing system based on deep potential function material simulation,and deep potential functions are trained in the system using neural networks that can be applied to a variety of atomic systems.Analyze the system’s ability to use heterogeneous computing resources,the efficiency and accuracy of the material simulation using deep potential functions are also investigated.From the results of this thesis,both of Docker and Singularity container technologies can approach physical machine in terms of performance of CPU and memory,but in the container batch processing scenario,both Docker Swarm and Kubernetes container orchestration services lack job management capabilities and container scheduling mechanisms are not perfect.In the container batch system based on classical potential function material simulation,Volcano can make up for the problems and defects of Kubernetes,and multiple containerized LAMMPS computing environments can also perform carbon material simulation well under classical potential functions such as AIREBO and Reax FF,and the simulation results of Reax FF potential function are better.Finally,in a container batch system based on deep potential function material simulation,Volcano can utilize computational resources more fully in deep learning training potential functions.Also,in experiments using GPU devices to accelerate the molecular dynamics simulation process,it was found that the deep potential function showed great advantages in computational efficiency and simulation accuracy in a wide range of atomic systems.Researchers in the materials field can perform materials simulations directly in the container batch system studied in this thesis,or they can customize the materials simulation environment according to the Dockerfile provided in this thesis,and the system can also be used by researchers in other fields. |
