CM-JP-11 Performance Tuning of Parallel Structural Analysis Code Based on Hierarchical Domain Decomposition Method for K Supercomputer
H Kawai, Masao Ogino, Ryuji Shioya, Sofia Lucas Yoshimura · The Proceedings of Mechanical Engineering Congress Japan · 2012
In this paper,a performance tuning approach of a structural analysis code based on hierarchical domain decomposition method(HDDM) for peta-scale massively parallel supercomputers is presented.Several types of subdomain local solver implementation are investigated.They are classified into four categories,DS(Direct solver-based matrix Storage),DSF(Direct Storage-Free),IS(Iterative solver-based matrix Storage) and ISF(Iterative Storage-Free).New local solvers,IS-ICT using Incomplete Cholesky factorization with Threshold pre-conditioner,and DS-LSC,using explicit evaluation of local Schur complement,are introduced.The implementation will be introduced to the future version of open-source CAE system,ADVENTURE.