Journal of Frontiers of Computer Science and Technology ›› 2014, Vol. 8 ›› Issue (8): 897-905.DOI: 10.3778/j.issn.1673-9418.1312035

Previous Articles     Next Articles

Accelerating First Principle Calculation Package on GPU Cluster

ZHAO Wei1,2, ZHAO Yonghua1+, LIU Xiaohui3, HE Lixin3   

  1. 1. Supercomputing Center, Computer Network Information Center, Chinese Academy of Sciences, Beijing 100190, China
    2. University of Chinese Academy of Sciences, Beijing 100049, China
    3. Key Lab of Quantum Information, University of Science and Technology of China, Hefei 230026, China
  • Online:2014-08-01 Published:2014-08-07


赵  慰1,2,赵永华1+,刘晓辉3,何力新3   

  1. 1. 中国科学院 计算机网络信息中心 超级计算中心,北京 100190
    2. 中国科学院大学,北京 100049
    3. 中国科学技术大学 量子信息重点实验室,合肥 230026

Abstract: First principle calculation package plays an important role in the development of density functional theory. Local orbital method is more suitable than plane wave method when calculating big system. Parallelism of the package becomes more and more important as the increasing of system scale and the development of HPC (high performance computing). The combination of MPI (message passing interface) and CUDA (compute unified device architecture) is a new way to accelerate the package. Accelerating the MESIA (massive electronic simulation based on systematically improvable atomic bases) package based on LCAO (linear combination of atomic orbitals) method with MPI+OpenMP+CUDA provides 15-fold speedup per GPU and good scalability. The test results also prove that computing with GPU can keep the accuracy of results.

Key words: compute unified device architecture (CUDA), grid integrals, first principle, local orbitals

摘要: 第一性原理计算软件在密度泛函理论的发展中起着重要作用。相比平面波方法,局域轨道法更适合处理大规模多体问题。随着问题规模的不断增大和计算机计算能力的提升,软件的并行加速成为一个重要课题,MPI(message passing interface)结合GPU(graphic processing unit)实现的异构并行是一个新的解决方案。基于局域轨道法的第一性原理计算软件MESIA(massive electronic simulation based on systematically improvable atomic bases)经过MPI+OpenMP+CUDA三级并行,单GPU取得了约15倍的加速比,同时表现出了良好的可扩展性。测试结果同时验证了使用GPU计算可以保证计算精度。

关键词: 统一计算架构(CUDA), 格点积分, 第一性原理, 局域轨道法