Add nospecialize to some very highly specialized methods

[odow]

x-ref #2829

[odow]

These cut 100k method instances. Still need to benchmark properly.

julia> length(after)
262910

[blegat]

This one may be more dangerous

[odow]

Note that the inner method will be compiled. It's just this one that isn't.

[blegat]

Maybe add a comment explaining this ?

[blegat]

Since it's a fallback returning false, I don't believe it will affect the performance of bottleneck to looks good

[blegat]

Maybe add a comment explaining this

[odow]

With lots of bridges

Slower because GC runs.

Before

julia> using JuMP, Gurobi, BenchmarkTools

julia> const ENV = Gurobi.Env()
Set parameter WLSAccessID
Set parameter WLSSecret
Set parameter LicenseID to value 722777
WLS license 722777 - registered to JuMP Development
Gurobi.Env(Ptr{Nothing} @0x000000012a809400, false, 0)

julia> function main(n)
           model = Model(() -> Gurobi.Optimizer(ENV))
           set_silent(model)
           @variable(model, x[1:n])
           @constraint(model, c[i in 1:n], 0 <= i * x[i] <= i)
           MOI.Utilities.attach_optimizer(model)
           return model
       end
main (generic function with 1 method)

julia> @benchmark(main(100_000); setup = GC.gc())
BenchmarkTools.Trial: 16 samples with 1 evaluation per sample.
 Range (min … max):  174.960 ms … 216.162 ms  ┊ GC (min … max): 0.00% … 18.86%
 Time  (median):     176.379 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   180.940 ms ±  12.739 ms  ┊ GC (mean ± σ):  2.60% ±  6.04%

   ██                                                            
  ▄██▄▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▄▁▁▁▁▁▁▁▄ ▁
  175 ms           Histogram: frequency by time          216 ms <

 Memory estimate: 344.67 MiB, allocs estimate: 6303323.

After

julia> using JuMP, Gurobi, BenchmarkTools

julia> const ENV = Gurobi.Env()
Set parameter WLSAccessID
Set parameter WLSSecret
Set parameter LicenseID to value 722777
WLS license 722777 - registered to JuMP Development
Gurobi.Env(Ptr{Nothing} @0x0000000127ddd000, false, 0)

julia> function main(n)
           model = Model(() -> Gurobi.Optimizer(ENV))
           set_silent(model)
           @variable(model, x[1:n])
           @constraint(model, c[i in 1:n], 0 <= i * x[i] <= i)
           MOI.Utilities.attach_optimizer(model)
           return model
       end
main (generic function with 1 method)

julia> @benchmark(main(100_000); setup = GC.gc())
BenchmarkTools.Trial: 14 samples with 1 evaluation per sample.
 Range (min … max):  213.427 ms … 235.628 ms  ┊ GC (min … max): 14.28% … 19.73%
 Time  (median):     229.635 ms               ┊ GC (median):    19.44%
 Time  (mean ± σ):   228.419 ms ±   5.341 ms  ┊ GC (mean ± σ):  18.89% ±  1.57%

  ▁                    ▁                 █▁ █   █ ██          ▁  
  █▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁█▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁██▁█▁▁▁█▁██▁▁▁▁▁▁▁▁▁▁█ ▁
  213 ms           Histogram: frequency by time          236 ms <

 Memory estimate: 349.26 MiB, allocs estimate: 6503956.

Without bridges

No change.

Before

julia> function main2(n)
           model = Model(() -> Gurobi.Optimizer(ENV))
           set_silent(model)
           @variable(model, x[1:n])
           @constraint(model, c[i in 1:n], i * x[i] >= i)
           MOI.Utilities.attach_optimizer(model)
           return model
       end
main2 (generic function with 1 method)

julia> @benchmark(main2(100_000); setup = GC.gc())
BenchmarkTools.Trial: 20 samples with 1 evaluation per sample.
 Range (min … max):  117.079 ms … 120.962 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     119.386 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   119.362 ms ± 868.501 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

  ▁                    ▁  █▁    ▁▁█  ▁ █  ▁▁ ▁▁   ▁ ▁    ▁    ▁  
  █▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁█▁▁██▁▁▁▁███▁▁█▁█▁▁██▁██▁▁▁█▁█▁▁▁▁█▁▁▁▁█ ▁
  117 ms           Histogram: frequency by time          121 ms <

 Memory estimate: 263.94 MiB, allocs estimate: 4802694.

After

julia> function main2(n)
           model = Model(() -> Gurobi.Optimizer(ENV))
           set_silent(model)
           @variable(model, x[1:n])
           @constraint(model, c[i in 1:n], i * x[i] >= i)
           MOI.Utilities.attach_optimizer(model)
           return model
       end
main2 (generic function with 1 method)

julia> @benchmark(main2(100_000); setup = GC.gc())
BenchmarkTools.Trial: 20 samples with 1 evaluation per sample.
 Range (min … max):  117.935 ms … 124.710 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     121.387 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   121.877 ms ±   1.833 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

                         ▃  ▃  █                              ▃  
  ▇▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▇▁▁▁▁▁▇█▁▁█▁▁█▁▁▇▇▁▇▁▁▁▁▁▁▁▁▇▁▁▁▁▁▇▁▇▇▁▁▁▁▁▇█ ▁
  118 ms           Histogram: frequency by time          125 ms <

 Memory estimate: 263.94 MiB, allocs estimate: 4802694.

[odow]

There doesn't seem to be any difference in the test runtime of master and this PR:
https://github.com/jump-dev/MathOptInterface.jl/actions/runs/17447865784/job/49546495222?pr=2830
https://github.com/jump-dev/MathOptInterface.jl/actions/runs/17505641833/job/49728541780

So maybe they're such tiny methods that it doesn't really matter.

[odow]

I guess I should try: https://timholy.github.io/SnoopCompile.jl/dev/tutorials/pgdsgui/. I remember looking at it a couple of years ago without much obvious success.

[odow]

So this has worked. I just needed to measure the first call, rather than subsequent. Because most of the benefit is avoiding the first compilation time.

Before

julia> using JuMP, Gurobi

julia> const ENV = Gurobi.Env()
Set parameter WLSAccessID
Set parameter WLSSecret
Set parameter LicenseID to value 722777
WLS license 722777 - registered to JuMP Development
Gurobi.Env(Ptr{Nothing} @0x0000000114808c00, false, 0)

julia> function main(n)
           model = Model(() -> Gurobi.Optimizer(ENV))
           set_silent(model)
           @variable(model, x[1:n])
           @constraint(model, c[i in 1:n], 0 <= i * x[i] <= i)
           MOI.Utilities.attach_optimizer(model)
           return model
       end
main (generic function with 1 method)

julia> @time main(100_000)
  1.185921 seconds (9.03 M allocations: 480.946 MiB, 13.67% gc time, 72.35% compilation time)
A JuMP Model
├ solver: Gurobi
├ objective_sense: FEASIBILITY_SENSE
├ num_variables: 100000
├ num_constraints: 100000
│ └ AffExpr in MOI.Interval{Float64}: 100000
└ Names registered in the model
  └ :c, :x

After

julia> using JuMP, Gurobi, BenchmarkTools

julia> const ENV = Gurobi.Env()
Set parameter WLSAccessID
Set parameter WLSSecret
Set parameter LicenseID to value 722777
WLS license 722777 - registered to JuMP Development
Gurobi.Env(Ptr{Nothing} @0x000000013f0d4c00, false, 0)

julia> function main(n)
           model = Model(() -> Gurobi.Optimizer(ENV))
           set_silent(model)
           @variable(model, x[1:n])
           @constraint(model, c[i in 1:n], 0 <= i * x[i] <= i)
           MOI.Utilities.attach_optimizer(model)
           return model
       end
main (generic function with 1 method)

julia> @time main(100_000)
  0.779390 seconds (8.10 M allocations: 427.501 MiB, 14.36% gc time, 61.28% compilation time)
A JuMP Model
├ solver: Gurobi
├ objective_sense: FEASIBILITY_SENSE
├ num_variables: 100000
├ num_constraints: 100000
│ └ AffExpr in MOI.Interval{Float64}: 100000
└ Names registered in the model
  └ :c, :x

Before

After

Code

using JuMP, Gurobi, BenchmarkTools
const ENV = Gurobi.Env()
function main(n)
    model = Model(() -> Gurobi.Optimizer(ENV))
    set_silent(model)
    @variable(model, x[1:n])
    @constraint(model, c[i in 1:n], 0 <= i * x[i] <= i)
    MOI.Utilities.attach_optimizer(model)
    return model
end
using SnoopCompileCore
tinf = @snoop_inference main(100_000);
using Profile
@profile main(100_000);
using SnoopCompile, PyPlot
mref, ax = pgdsgui(tinf);

[odow]

Even without bridges this is much better.

Before

julia> using JuMP, Gurobi

julia> const ENV = Gurobi.Env()
Set parameter WLSAccessID
Set parameter WLSSecret
Set parameter LicenseID to value 722777
WLS license 722777 - registered to JuMP Development
Gurobi.Env(Ptr{Nothing} @0x0000000142a8ac00, false, 0)

julia> function main2(n)
           model = Model(() -> Gurobi.Optimizer(ENV))
           set_silent(model)
           @variable(model, x[1:n])
           @constraint(model, c[i in 1:n], i * x[i] >= i)
           MOI.Utilities.attach_optimizer(model)
           return model
       end
main2 (generic function with 1 method)

julia> @time main2(100_000)
  0.338462 seconds (4.99 M allocations: 273.114 MiB, 26.99% gc time, 38.53% compilation time)
A JuMP Model
├ solver: Gurobi
├ objective_sense: FEASIBILITY_SENSE
├ num_variables: 100000
├ num_constraints: 100000
│ └ AffExpr in MOI.GreaterThan{Float64}: 100000
└ Names registered in the model
  └ :c, :x

After

julia> using JuMP, Gurobi

julia> const ENV = Gurobi.Env()
Set parameter WLSAccessID
Set parameter WLSSecret
Set parameter LicenseID to value 722777
WLS license 722777 - registered to JuMP Development
Gurobi.Env(Ptr{Nothing} @0x0000000118251000, false, 0)

julia> function main2(n)
           model = Model(() -> Gurobi.Optimizer(ENV))
           set_silent(model)
           @variable(model, x[1:n])
           @constraint(model, c[i in 1:n], i * x[i] >= i)
           MOI.Utilities.attach_optimizer(model)
           return model
       end
main2 (generic function with 1 method)

julia> @time main2(100_000)
  0.228506 seconds (4.84 M allocations: 265.752 MiB, 42.85% gc time, 6.42% compilation time)
A JuMP Model
├ solver: Gurobi
├ objective_sense: FEASIBILITY_SENSE
├ num_variables: 100000
├ num_constraints: 100000
│ └ AffExpr in MOI.GreaterThan{Float64}: 100000
└ Names registered in the model
  └ :c, :x

[odow]

Thoughts @blegat?

[odow]

Bump @blegat

[joaquimg]

Perhaps a question for a separate PR, would it make sense to not-specialize most of the functions in MOI.Test?

[blegat]

Looks good to me, just need to address https://github.com/jump-dev/MathOptInterface.jl/pull/2830/files#r2378993412

[odow]

I reverted the add_constraint changes. We don't have to go overboard. Some are better than none.

[odow]

Before

(grb2) pkg> st
Status `/private/tmp/grb2/Project.toml`
  [2e9cd046] Gurobi v1.7.6
  [4076af6c] JuMP v1.29.1

julia> using JuMP, Gurobi

julia> const ENV = Gurobi.Env()
Set parameter WLSAccessID
Set parameter WLSSecret
Set parameter LicenseID to value 722777
WLS license 722777 - registered to JuMP Development
Gurobi.Env(Ptr{Nothing} @0x000000014db87200, false, 0)

julia> function main(n)
           model = Model(() -> Gurobi.Optimizer(ENV))
           set_silent(model)
           @variable(model, x[1:n])
           @constraint(model, c[i in 1:n], 0 <= i * x[i] <= i)
           MOI.Utilities.attach_optimizer(model)
           return model
       end
main (generic function with 1 method)

julia> @time main(100_000)
  1.099357 seconds (8.99 M allocations: 479.260 MiB, 12.73% gc time, 71.90% compilation time)
A JuMP Model
├ solver: Gurobi
├ objective_sense: FEASIBILITY_SENSE
├ num_variables: 100000
├ num_constraints: 100000
│ └ AffExpr in MOI.Interval{Float64}: 100000
└ Names registered in the model
  └ :c, :x

After

(grb) pkg> st
Status `/private/tmp/grb/Project.toml`
  [2e9cd046] Gurobi v1.7.6
  [4076af6c] JuMP v1.29.1
  [b8f27783] MathOptInterface v1.45.0 `~/git/jump-dev/MathOptInterface`

julia> using JuMP, Gurobi

julia> const ENV = Gurobi.Env()
Set parameter WLSAccessID
Set parameter WLSSecret
Set parameter LicenseID to value 722777
WLS license 722777 - registered to JuMP Development
Gurobi.Env(Ptr{Nothing} @0x000000014f1cb000, false, 0)

julia> function main(n)
           model = Model(() -> Gurobi.Optimizer(ENV))
           set_silent(model)
           @variable(model, x[1:n])
           @constraint(model, c[i in 1:n], 0 <= i * x[i] <= i)
           MOI.Utilities.attach_optimizer(model)
           return model
       end
main (generic function with 1 method)

julia> @time main(100_000)
  0.750228 seconds (7.83 M allocations: 413.546 MiB, 17.80% gc time, 57.17% compilation time)
A JuMP Model
├ solver: Gurobi
├ objective_sense: FEASIBILITY_SENSE
├ num_variables: 100000
├ num_constraints: 100000
│ └ AffExpr in MOI.Interval{Float64}: 100000
└ Names registered in the model
  └ :c, :x