Top Page > Browsing

One more question about NEGF performance Date: 2014/08/13 18:38 Name: Artem   <artem.pulkin@epfl.ch> Dear Taisuke, I collected time statistic of NEGF iteration. Specifically, I calculated several systems with the same leads but scattering regions were of different size. Each calculation was performed by mpirun -np 48 on a 48-core node with 192G of memory. The time refers to <TRAN_DFT> time=[*]. Atoms (total with leads) | Hamiltonian size | time (s) 34 | 1304 | 487 47 | 1792 | 1019 68 | 2572 | 2626 71 | 2704 | 3189 80 | 3028 | 14388 82 | 3128 | 16033 83 | 3160 | 16039 It looks like there is a large gap between 71 and 80 atoms calculation. Do you know the reason for that? According to "top" there is still at least 50G free on each node. My makefile: CC = mpiicc -openmp -O3 -I/opt/intel/12.1.2/mkl/include/fftw/ FC = mpiifort -openmp -O3 -I/opt/intel/12.1.2/mkl/include/fftw/ LIB = -lmpiif -L/opt/intel/12.1.2/mkl/lib/intel64/ -lmkl_intel_lp64 -lmkl_sequential -lmkl_core -lifcore Regards, Artem

Page: [1]

Re: One more question about NEGF performance ( No.1 ) Date: 2014/08/18 18:17 Name: Artem  <artem.pulkin@epfl.ch> This one solved: I switched to two nodes and 96 cores and TRAN_DFT reduced by a factor of 5. Probably the reason was overcommunication in LAPACK.

Page: [1]