arfc / publications

public repository for the development and distribution of publications
Creative Commons Attribution 4.0 International
2 stars 7 forks source link

Moltres init: In-depth scaling tests #42

Closed gridley closed 7 years ago

gridley commented 7 years ago

The proposed scaling test is very limited. Proper scaling tests would require weak and strong scaling tests, an assessment on the impact of the size of problem per core, a separate assessment of intra-node and extra-node scaling, etc. In addition, information should be provided on the employed numerical algorithms and on the parallelization strategy. The authors should either extend significantly the section, or remove it completely.

Maybe other MOOSE results could presented, and then present some Moltres results in the same terms. This would show that the solvers and parallelization provided by MOOSE work well for the physics exhibited by MSRs.

andrewryh commented 7 years ago

I can do strong and weak scaling study for moltres on BW, learning it right now in my Parallel Algos class. If interested, pick your favorite case which is appropriate for the study.

lindsayad commented 7 years ago

How big should the problem be?

On Tue, Oct 3, 2017 at 12:06 PM, Andrei Rykhlevskii < notifications@github.com> wrote:

I can do strong and weak scaling study for moltres on BW, learning it right now in my Parallel Algos class. If interested, pick your favorite case which is appropriate for the study.

— You are receiving this because you were assigned. Reply to this email directly, view it on GitHub https://github.com/arfc/publications/issues/42#issuecomment-333929477, or mute the thread https://github.com/notifications/unsubscribe-auth/AJxgcKTLPF0S6xa9mqiYrJXMe6R1LoRTks5sonepgaJpZM4Pn57p .

andrewryh commented 7 years ago

It depends on how deep we want to study that. You have mentioned 608 cores in an introduction, but I think 320 cores (10 nodes) quite enough. So, pick a problem which requires ~2-10 minutes on 320 cores.

For weak scaling we have to fix a number of points per processor, so, I will scale mesh while changing core number, hope I know how to do it in gmsh.

@katyhuff Don't you mind if I'd take this assignment?

katyhuff commented 7 years ago

@andrewryh Sounds good to me, but it's up to alex. I think he thought this topic was interesting and self-assigned it .

lindsayad commented 7 years ago

You go for it Andrei. I'll look at concocting a problem then and pass it onto you

On Oct 3, 2017, at 3:54 PM, Katy Huff notifications@github.com wrote:

@andrewryh Sounds good to me, but it's up to alex. I think he thought this topic was interesting and self-assigned it .

— You are receiving this because you were assigned. Reply to this email directly, view it on GitHub, or mute the thread.

lindsayad commented 7 years ago

@andrewryh I think I would run moltres/problems/publication_level_cases/3d_steady_state/3d_auto_diff_rho.i and change end_time to something like 1 second. You might have to make it even smaller to get a simulation time in the minute to ten minutes range.

andrewryh commented 7 years ago

Gotcha. I'll change somehow mesh to variate number of grid point per processor also. Thank, don't forget assign it to me :)

katyhuff commented 7 years ago

Closed with #58