Pressure solver performance for PNPN about nek5000 HOT 5 CLOSED

@ggiannako If I recall correctly, you did performance experiments. Can you add more details?

from nek5000.

In a small 3D engine sim case (E=75k, lx1=6), for 100 steps (fixed dt), the average number of pressure iterations for PN/PN is 36% higher than PN/PN-2. In this case, this results to about 23% higher runtime.

from nek5000.

from nek5000.

Hi,

I have some data from a channel flow comparison we did quite some time ago, so not sure how relevant it would be. It is from 2014, and was done with version 758. We have three relevant cases:

1. 97,344 elements, N=11, Pn-Pn-2. Time per time step: 8.05e-1

2. 97,344 elements, N=11, Pn-Pn. Time per time step: 8.18e-1 [around 2% higher]

3. 97,344 elements, N=11, Pn-Pn + characteristics. Time per time step: 2.74, for an increase in time step of a factor of 2. So if one would multiply the value 8.18e-1 by 2 one would get 1.63, still much lower than the value with characteristics.

I think that at the time we concluded that the performance was similar between Pn-Pn-2 and Pn-Pn (perhaps in some runs the difference was a bit more than 2%, but not so much).

Best regards,

Ricardo.

from nek5000.

Just checked turbChannel with different meshes, XXT and AMG. The difference in runtime between Pn-Pn and Pn-Pn-2 was lower than 10%.

For now, I am closing this issue as no longer reproducible.

from nek5000.