[FileFormats.MOF] replace OrderedDict by NamedTuple when writing

[joaquimg]

All OrderedDicts used are very small. Hence, they allocate unnecessarily. This also creates more GC pressure.

[odow]

Could we get some benchmarks?

[odow]

This is quite nice

[odow]

I opened an issue in JuliaFormatter to discuss formatting the named tuples: domluna/JuliaFormatter.jl#895

[joaquimg]

A test

using JuMP

model = Model()
@variable(model, x[1:1000] >= 0)
@constraint(model, y[1:1000], sum(i*x[i] for i in 1:1000) == 1)

for i in 1:5
    GC.gc()
    @time write_to_file(model, "test.mof.json")
end

Results:

before PR:

  5.528170 seconds (20.16 M allocations: 1.669 GiB, 31.54% gc time)
  5.244265 seconds (20.16 M allocations: 1.669 GiB, 30.18% gc time)
  5.487519 seconds (20.16 M allocations: 1.669 GiB, 29.10% gc time)
  5.344851 seconds (20.16 M allocations: 1.669 GiB, 30.58% gc time)
  5.348303 seconds (20.16 M allocations: 1.669 GiB, 29.13% gc time)

after PR:

  2.293431 seconds (5.13 M allocations: 936.505 MiB, 13.93% gc time)
  2.281071 seconds (5.13 M allocations: 936.508 MiB, 13.94% gc time)
  2.995470 seconds (5.13 M allocations: 936.454 MiB, 12.53% gc time)
  3.088288 seconds (5.13 M allocations: 936.497 MiB, 12.26% gc time)
  2.971376 seconds (5.13 M allocations: 936.498 MiB, 12.24% gc time)

Which is a big win as there is also the fixed cost of writing the json that both incur.

[odow]

Nice nice nice.

Note that this could be regarded as technically breaking, but I don't think anyone is using moi_to_object: https://juliahub.com/ui/Search?type=code&q=moi_to_object. It's also not mentioned in the docs, so I'm going to say that it is the private API.

[odow]

Reduction in memory pressure is a big deal for the larger Sienna models.

cc @jd-lara this should help your 'always write to MOF'

[joaquimg]

should we rename moi_to_object to _moi_to_object?

[odow]

Nah leave it for now

…

On Fri, 10 Jan 2025, 10:34 am Joaquim Dias Garcia, ***@***.***> wrote: should we rename moi_to_object to _moi_to_object? — Reply to this email directly, view it on GitHub <#2606 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AB6MQJOZMWCC44ZMIHOMWH32J3TOVAVCNFSM6AAAAABU466FA6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKOBRGI4DQNRUGY> . You are receiving this because you commented.Message ID: ***@***.***>

[joaquimg]

green light from CI

[odow]

This should make such a difference; we were making a new OrderedDict for every single coefficient in the model 😭

[odow]

@joaquimg do you have a script for the before/after Sienna profile flame graph?

[joaquimg]

Is the same script used here (but with PProf):

NREL-Sienna/PowerSimulations.jl#1183