Parallel model building docs

[joaquimg]

Users keep asking about it and some are already doing it with @expression and @build_constraint.

We should add comments about it in the parallelism tutorial.

[odow]

We already have
https://jump.dev/JuMP.jl/stable/tutorials/algorithms/parallelism/#When-building-a-JuMP-model

Does building in parallel really work/improve things?

[odow]

Closing in favor of #2348 (comment)

[odow]

I've just never seen a case where this is actually meaningful.

julia> using JuMP

julia> function build_threaded(N)
           model = Model()
           @variable(model, x[1:N])
           my_lock = Threads.ReentrantLock();
           Threads.@threads for i in 1:N
               con = @build_constraint(sum(x[1:i]) >= 0)
               lock(my_lock) do
                   add_constraint(model, con)
               end
           end
           return model
       end
build_threaded (generic function with 1 method)

julia> function build_serial(N)
           model = Model()
           @variable(model, x[1:N])
           for i in 1:N
               @constraint(model, sum(x[1:i]) >= 0)
           end
           return model
       end
build_serial (generic function with 1 method)

julia> @time build_serial(1_000);
  0.100237 seconds (47.79 k allocations: 108.902 MiB)

julia> @time build_serial(1_000);
  0.085962 seconds (47.79 k allocations: 108.902 MiB)

julia> @time build_threaded(1_000);
  0.067640 seconds (71.92 k allocations: 110.531 MiB, 247.57% compilation time)

julia> @time build_threaded(1_000);
  0.038823 seconds (47.82 k allocations: 108.904 MiB)

julia> @time build_threaded(10_000);
 14.665181 seconds (642.39 k allocations: 10.029 GiB, 42.87% gc time)

julia> @time build_serial(10_000);
 19.576950 seconds (642.36 k allocations: 10.029 GiB, 14.43% gc time)

julia> Threads.nthreads()
5

[jd-lara]

We use parallelism to build the expressions used for the network flow constraints. We pre-generate all the expression in parallel doing the vector product of the variables and the column vector of the PTDF to generate the affine expressions and THEN we post the constraints.

That being said, it took a while to get it right.

https://github.com/NREL-Sienna/PowerSimulations.jl/blob/cfc1f52411797528d4d090a866cda3bb5f0cdfb5/src/devices_models/devices/AC_branches.jl#L617

[odow]

Does it make a difference? Do you have benchmarks?

[jd-lara]

Does it make a difference? Do you have benchmarks?

It makes a difference for super-large networks and doesn't hurt for smaller ones. I could test again, the serial code is still there for benchmarking and debugging.

[odow]

Okay. If people are going to do this, we should at least give instructions for how to do it safely

[odow]

Closing again in favor of #2348 (comment)