getting process layout for rectangular domains?

[edoyango]

I need to generate domains for CPU v GPU perf plots for MOM6. I was going to do square domains, but rectangular process layouts fit better on the number of CPUs I'm testing. Atm I'm testing 18 and 104 cores - the best layouts being 6x3 and 13x8 respectively. I thought it would be useful to test both square and rectangular domains to see how the GPU version performs.

I'm happy to implement this feature but wanted input on where to implement it. In find_layouts_with_maxncore? e.g. with an arg like rectangle_ratio? 1 would give square, 2 would try to give processor layout of 2:1 etc.

[manodeep]

Square is covered - if you set abs_maxdiff_nx_ny to 0, then you will only get back layouts with nx==ny.

To confirm, you want (nx, ny) tuples that satisfy (nx * ny) < maxncores && (nx/ny >= min_ratio && nx/ny <= max_ratio) ?

[edoyango]

yes something like that. I'd prefer not to have to specify a min and max ratio though. Perhaps could use the abs_maxdiff arg?

but instead of the diff being between nx and ny, it would be nx and nx_ideal, and ny and ny_ideal e.g. if my ideal was 6 and 3, and my abs_maxdiff was 1, then 5x3, 7x2 all would be ok, but 4x4 wouldn't.

[manodeep]

Currently, the ideal is assumed to be square, and effectively both nx_ideal and ny_ideal are set to $\sqrt{{\rm maxncore}}$

(side-note: GH markdown has good LaTeX support)

[edoyango]

so i'm hoping for something like

>>> find_layouts_with_maxncore(18, nx_ny_fraction=2, abs_maxdiff=1)
[
    (6, 3),
    (5, 3),
    (7, 2),
    # etc.
]

probably not the best example, but the tuple should be around (6, 3) and if nx_ny_fraction = 0.5 then they should be around (3, 6).

[edoyango]

I've realized the functionality I'm after is a bit different to what the existing find_layouts does.

Find_layout asks "what process layout should I have if I have n cores?" Whereas I'm asking "what is my domain shape if I have nx and ny processes?" Since in my case, I have a fixed number of processes, but more flexibility in domain shape.

I reckon I'll implement this in a different function. Any opinions?

[manodeep]

Don't think I follow the requirement - what can domain shape be, and how is it related to the processor layout?

[micaeljtoliveira]

I think all of these cases can be formulated generally like this: given an horizontal grid, a total number of cores, and some user-defined constrains, what are the valid (nx, ny) tuples to partition the grid.

Whereas I'm asking "what is my domain shape if I have nx and ny processes?"

That's not very complicated, you just need to know the number of grid points along each direction. The problem is that not all choices of nx, ny are valid, so before you ask your question, you need to get a list of valid tuples.

I reckon I'll implement this in a different function. Any opinions?

I would keep the workflow the way you described it first: get a list of valid tuples, then choose one from that list for your run.