When using VASP with multiple nodes, User time and System time are too long

[Jike.Wang]

Dear Admin and Users,

Now I am trying to use VASP with two nodes, but there is a problem with the running time. A system containing 32 atoms is used for testing, and the running time of different number of cores and nodes is as follows:


User time (sec) System time (sec) Total CPU time used (sec) Elapsed time (sec)
16cores 1node 1631.324 3.028 1634.352 1769.645

32cores 1node 1062.693 2.743 1065.436 1141.688

32cores 2nodes 1850.855 1101.752 2953.607 3097.869


When using two nodes, both User time and System time become longer, which reduces computational efficiency. Therefore, I have some questions,

1. What factors affect User time and System time, respectively?
2. How to shorten User time and System time?
3. Is there a large difference in computational efficiency between ten-gigabit switch and Infiniband? How much?

Thank you for taking the time to focus on my problems!

[fabien_tran1]

Hi,

In case you have not yet searched for information about parallelism with VASP, then this should be helpful:
https://www.vasp.at/wiki/index.php/Cate ... lelization
https://www.youtube.com/watch?v=KzIuL_e0zz8