5:24:25 AM PDT - Mon, Jun 17th 2019 |
|
RT TDDFT serial and parallel performance
|
I ran a test using the water excitation example posted in the documentation. The serial results agree with the expected output for total energy, dipole moment, electric field and statistics on the data.
Now when I use 4 cores with mpi the execution time is twice as long and the results are different so that plots of the time evolution of any of the quantities mentioned are very different.
serial parallel reference std. dev (serial, parallel, ref)
total energy -76.301190 -76.30120 -76.301190 (0.000005, 7.195892e-09, 0.000005)
dipole moment -0.984565 -0.984648 -0.984565 (0.003478, 0.000107, 0.003478)
electric field 2.505774e-15 -1.242907e-05 2.505774e-15 (2.396883e-05, 2.275951e-05, 2.396883e-05)
The issue is not the mean values but the time evolution of the system with the same initial state and the poor performance when multiple cores are used on a single job.
The electric does not even show an oscillating amplitude when run as a parallel job. Any suggestions for the behavior differences?
|