Node mapping and the parallel performance of the LS3DF code on BG/P

Zhengji Zhao, NERSC User Services Group, Lawrence Berkeley National Laboratory

The Linearly scaling 3 dimensional fragment (LS3DF) code for large scale ab initio electronic structure calculations scales almost perfectly to the full machine size (163,8 40 cores) on BG/P. It was not possible to achieve such a parallel scaling without a careful design of the physical lay out (mapping) of the mpi tasks. In this talk we will present how the node mapping affect the parallel performance of the LS3DF code, as well as the parallel performance of the LS3DF code on BG/P.

ScicomP logo