The high-throughput Computing Algorithm and Data Storage of First-principle Calculation Based on the Optimized Domestic x86 Chip

Chaofang Dong¹*, Yucheng Ji¹, Song Wang²*, Xiguo Xie², Dihao Chen¹, Ni Li¹, Xiaogang Li¹

¹ Beijing Advanced Innovation Center for Materials Genome Engineering, National Material Environmental Corrosion Science Data Center, Institute for Advanced Materials and Technology, University of Science and Technology Beijing, Beijing 100083, China

² Dawning Information Industry (Beijing) Co., Ltd. Beijing 100193, China

ABSTRACT: To stabilize material batch fluctuations and promote the further improvement of material performance for major projects and equipment, it is necessary to control the composition and structure on the micro-scale inside the material. Under the concept of Material Genetic Engineering (MGE), a data chain construction of data accumulation, mining modeling, simulation and sharing has been carried out, hoping to accelerate the development of new materials.

In the material simulation, high-throughput data accumulation and high-quality big data storage are important guarantees for improving the accuracy and application scope of performance evaluation models. Owing to the key role of high-performance multi-core processors in the computing data acquisition, the collaborative optimization of software and hardware with the processor has become constraints in domestic high-performance computing industry. With the development of supercomputing center in Beijing, Tianjin, Shanxi, Jiangsu, Shenzhen, Chengdu, Jinan and other places, the scale and influence of high-performance computing industry continue to expand. However, the supply of key chips restricts the rapid development of high-performance computing, it is urgent to develop a new generation of domestic x86 chips which are optimized for material calculations.

The mainstream methods of material simulation are still first-principles, molecular dynamics and finite element methods. First-principle calculation is a professional approach for studying the structure and electronic properties of materials, which size of a single submission is usually limited to ten or a hundred. This traditional task submission method limits the collection of large-scale data and the high computing performance of domestic chips. Therefore, this work has developed a high-throughput computing algorithms based on x86 chips for first-principle calculation. The compiled high-throughput calculation and data storage code can complete the large-scale data generation, submission, calculation, monitoring, and result processing. The scale of a single submission has been increased to 1,000, and materials data specifications have been uniformly formulated to strength data accuracy.

Based on the above-mentioned high-throughput algorithm which integrated first-principle calculation and data automatic storage, this study forces on the corrosion resistant design of Al alloys for high-speed trains and develops a material structure-performance data collection and data fusion platform for MGE. National encryption algorithm (SM1, SM2, SM4) and VASP optimization are applied to the domestic x86 chip, which not only improves the efficiency of high-throughput data acquisition of material calculation, but also ensures the data security. Firstly, the high-throughput algorithm is compiled based on the domestic x86 chip platform, the BLIS and FFTW libraries used by VASP have been professionally optimized. After testing and verification, the computational efficiency after optimized is 1.5 to 2 times that of other supercomputing platforms with the same resources. Next, the algorithm intelligently generates and optimizes the calculation file according to the sites, doping elements and the basic models. During the calculation, the monitoring program dynamically retrieves the task status and determines whether the calculation is correct. After the calculation is completed and correct, the algorithm invokes the extraction program to automatically complete the data storage. Finally, the algorithm embeds computing formulas, such as grain boundary cohesive energy, to improve the efficiency of data processing. By calling the results in the database, the final data processing is completed at one time. With the aid of this algorithms, the potential alloy elements were screened by this high-through computing algorithm to lower the susceptibility of the weak GBs to the intergranular cracking that is caused by stress corrosion cracking.

Keywords: High-throughput algorithm; Database; First-principle calculation; Domestic x86 chips.