Informatické kolokvium 17. 5. Highly scalable domain decomposition methods on many-core . . .
Informatické kolokvium 17. 5. 2016, 14:00 posluchárna D2
doc. Mgr. Vít Vondrák, Ph.D., IT4Innovations, VŠB-TUO
Highly scalable domain decomposition methods on many-core architectures
Abstrakt: The latest IT4Innovations computing system - Salomon includes significant number of many integrated cores architecture. An efficient use of such architecture needs reviewing of current algorithms and their adapting for such system. The research team at IT4Innovations is focused on development of highly scalable algorithms for solution of linear and non-linear problems by a non-overlapping domain decompositions method called Finite Element Tearing and Interconnecting method. The most efficient variants were adapted to the new MIC architecture and implemented into our in-house software package ESPRESO (ExaScale PaRallel FETI Solver). Its developers focus on highly efficient communication layer on top of pure MPI and efficient use of MIC architectures to improve solvers performance and scalability. Thus ESPRESO uses three levels of parallelization: 1.) message passing on node level, 2) threading inside a node using Cilk++ and 3.) vectorization using Intel MKL library or Cilk++. The experience with running of ESPRESO solvers on Salomon and other world’s top computing systems will be presented.
doc. Mgr. Vít Vondrák, Ph.D., IT4Innovations, VŠB-TUO
Highly scalable domain decomposition methods on many-core architectures
Abstrakt: The latest IT4Innovations computing system - Salomon includes significant number of many integrated cores architecture. An efficient use of such architecture needs reviewing of current algorithms and their adapting for such system. The research team at IT4Innovations is focused on development of highly scalable algorithms for solution of linear and non-linear problems by a non-overlapping domain decompositions method called Finite Element Tearing and Interconnecting method. The most efficient variants were adapted to the new MIC architecture and implemented into our in-house software package ESPRESO (ExaScale PaRallel FETI Solver). Its developers focus on highly efficient communication layer on top of pure MPI and efficient use of MIC architectures to improve solvers performance and scalability. Thus ESPRESO uses three levels of parallelization: 1.) message passing on node level, 2) threading inside a node using Cilk++ and 3.) vectorization using Intel MKL library or Cilk++. The experience with running of ESPRESO solvers on Salomon and other world’s top computing systems will be presented.