Scalability Study, Keeping Total Grid Cells Constant Across Mesh Configurations

[MechAustin]

Hi all,
I wanted to open a discussion on how I should approach the scalability study while keeping the total number of grid cells the same across different processor configurations.
My plan is to use a base grid of 28×32×21 cells for the 1×1 (single process) case, and then subdivide that same total grid across more processors without changing the overall problem size. Here is how I broke it down:

1×1 → 28×32×21 cells total (one mesh/process)
2×2 → 14×16×21 cells per mesh/process (4 processes total)
4×4 → 7×8×21 cells per mesh/process (16 processes total)

My goal is to test strong scaling — I want to keep the total work fixed and measure how much faster it runs as I add more processors.
One issue I ran into is that when I subdivide the domain for the 4×4 case, I am getting 75,264 total grid cells, which does not match my 1×1 case. I am wondering if ghost cells or boundary overlaps between subdomains are causing this difference.
Can anyone help clarify if my approach is correct and what might be causing that discrepancy?
Thanks
CamScanner 5-7-26 18.27.pdf

[drjfloyd]

See 3.3.1 in the User's Guide. The input files can be found in:

https://github.com/firemodels/fds/tree/master/Validation/MPI_Scaling_Tests/FDS_Input_Files

The weak scaling inputs keep the compute per process the same and the strong scaling applies more and more processes to the same compute.

You should expect to see a few stages representing before and after hitting the limits of your memory bus and your network.