Use MOI.add_constrained_variable when creating variables

[odow]

Replacement for #3862

See also jump-dev/HiGHS.jl#245

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 97.55%. Comparing base (c63b19c) to head (a79859c).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #3863      +/-   ##
==========================================
+ Coverage   97.54%   97.55%   +0.01%     
==========================================
  Files          44       44              
  Lines        6029     6059      +30     
==========================================
+ Hits         5881     5911      +30     
  Misses        148      148              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[odow]

Before

julia> using JuMP, HiGHS

julia> function main()
           model = direct_model(HiGHS.Optimizer())
           set_silent(model)
           set_string_names_on_creation(model, false)
           @variable(model, 1 <= x[i in 1:10_000] <= 2)
           return model
       end
main (generic function with 1 method)

julia> @time main();
  0.025593 seconds (30.06 k allocations: 2.307 MiB)

julia> using BenchmarkTools

julia> @benchmark main()
BenchmarkTools.Trial: 198 samples with 1 evaluation.
 Range (min … max):  24.018 ms …  31.237 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     25.140 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   25.339 ms ± 981.654 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

       ▁▂▂▄ ▁▂█▇▆▄                                              
  ▃▁▁▆▆████▅████████▃▄▁▃▆▁▃▁▃▆▁▁▁▃▃▃▃▃▃▁▁▃▁▁▁▁▁▃▁▃▁▁▃▁▁▁▁▁▁▃▁▃ ▃
  24 ms           Histogram: frequency by time         29.2 ms <

 Memory estimate: 2.31 MiB, allocs estimate: 30056.

(hgh) pkg> st
Status `/private/tmp/hgh/Project.toml`
  [87dc4568] HiGHS v1.10.2
  [4076af6c] JuMP v1.23.3

After

julia> using JuMP, HiGHS

julia> function main()
           model = direct_model(HiGHS.Optimizer())
           set_silent(model)
           set_string_names_on_creation(model, false)
           @variable(model, 1 <= x[i in 1:10_000] <= 2)
           return model
       end
main (generic function with 1 method)

julia> @time main();
  0.016782 seconds (30.06 k allocations: 2.307 MiB)

julia> using BenchmarkTools

julia> @benchmark main()
BenchmarkTools.Trial: 284 samples with 1 evaluation.
 Range (min … max):  16.344 ms … 29.127 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     17.092 ms              ┊ GC (median):    0.00%
 Time  (mean ± σ):   17.609 ms ±  1.672 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

  ▄█▆▆▆▆▄▄▃▁▂ ▁                                                
  ██████████████▆█▆▆▆▁▄▆▆▆▆▄▄▁▄▄▄▁▁▁▁▆▄▁▄▄▁▁▄▁▁▁▁▁▄▁▄▁▁▁▁▁▁▁▄ ▆
  16.3 ms      Histogram: log(frequency) by time      25.5 ms <

 Memory estimate: 2.31 MiB, allocs estimate: 30056.

(hgh2) pkg> st
Status `/private/tmp/hgh2/Project.toml`
  [87dc4568] HiGHS v1.10.2 `https://github.com/jump-dev/HiGHS.jl.git#master`
  [4076af6c] JuMP v1.23.3 `https://github.com/jump-dev/JuMP.jl.git#od/add-constrained-variable`

Change

So from 24ms to 16ms, which is about 1/3 faster, or exactly what we predicted might be the case on the call.

Single-sided bounds

Single-side bounds are twice as fast. From 16ms to 8ms:

julia> using JuMP, HiGHS

julia> function main()
           model = direct_model(HiGHS.Optimizer())
           set_silent(model)
           set_string_names_on_creation(model, false)
           @variable(model, x[i in 1:10_000] >= 0)
           return model
       end
main (generic function with 1 method)

julia> @time main();
  0.049627 seconds (41.81 k allocations: 3.119 MiB, 62.90% compilation time)

julia> using BenchmarkTools

julia> @benchmark main()
BenchmarkTools.Trial: 274 samples with 1 evaluation.
 Range (min … max):  16.755 ms …  22.301 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     18.170 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   18.237 ms ± 583.379 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

                      ▂  ▄▁█▁▃▄▆▄▃▂ ▄▁▃   ▁▂                    
  ▄▁▁▁▁▁▁▁▃▄▁▄▃▅▅▆▇█▆▅█▇▇██████████▇████▅▅██▆▆▇▃▃▅▃▃▄▃▄▄▄▁▄▁▁▄ ▄
  16.8 ms         Histogram: frequency by time         19.6 ms <

 Memory estimate: 2.31 MiB, allocs estimate: 30056.

to

julia> using JuMP, HiGHS

julia> function main()
           model = direct_model(HiGHS.Optimizer())
           set_silent(model)
           set_string_names_on_creation(model, false)
           @variable(model, x[i in 1:10_000] >= 0)
           return model
       end
main (generic function with 1 method)

julia> @time main();
  0.041924 seconds (41.62 k allocations: 3.094 MiB, 73.63% compilation time)

julia> using BenchmarkTools

julia> @benchmark main()
BenchmarkTools.Trial: 468 samples with 1 evaluation.
 Range (min … max):   8.908 ms … 375.817 ms  ┊ GC (min … max): 0.00% … 93.33%
 Time  (median):      9.863 ms               ┊ GC (median):    0.00%
 Time  (mean ± σ):   10.672 ms ±  16.926 ms  ┊ GC (mean ± σ):  7.02% ±  4.31%

     ▃▁█▃▅▆  ▁  ▂ ▁▃ ▄ ▆▁▇▃ ▂▄▁▃▃▁                              
  ▃▃▆██████▅██▇▇█▇████▇████▅██████▅▄▆▆▃▇▆▄▄▃▃▃▄▃▄▁▃▁▃▃▃▃▃▁▃▃▁▃ ▄
  8.91 ms         Histogram: frequency by time         11.7 ms <

 Memory estimate: 2.31 MiB, allocs estimate: 30056.

[odow]

So this is an initial backwards compatible minimal change that is a good performance improvement, but only partially solves the problem.

@joaquimg, you could always write your variables as x in MOI.Interval(l, u) to get another speedup from

julia> function main()
           model = direct_model(HiGHS.Optimizer())
           set_silent(model)
           set_string_names_on_creation(model, false)
           @variable(model, x[i in 1:10_000] in MOI.Interval(0.0, 1.0))
           return model
       end
main (generic function with 1 method)

julia> @time main();
  0.098376 seconds (82.31 k allocations: 7.882 MiB, 63.35% compilation time)

julia> using BenchmarkTools

julia> @benchmark main()
BenchmarkTools.Trial: 156 samples with 1 evaluation.
 Range (min … max):  30.089 ms … 36.249 ms  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     31.996 ms              ┊ GC (median):    0.00%
 Time  (mean ± σ):   32.061 ms ±  1.038 ms  ┊ GC (mean ± σ):  0.00% ± 0.00%

           ▃  ▃▃▅▅ ▂ ▃▆█ ▆ ▆▃                                  
  ▄▄▅▅▄▇▇▅▄█▅███████████▅████▇██▇▇█▅▅▅▄▁▄▁▄▄▁▁▄▄▁▁▁▁▄▁▄▁▁▁▁▁▄ ▄
  30.1 ms         Histogram: frequency by time        35.5 ms <

to

julia> using JuMP, HiGHS

julia> function main()
           model = direct_model(HiGHS.Optimizer())
           set_silent(model)
           set_string_names_on_creation(model, false)
           @variable(model, x[i in 1:10_000] in MOI.Interval(0.0, 1.0))
           return model
       end
main (generic function with 1 method)

julia> @time main();
  0.087588 seconds (67.48 k allocations: 6.796 MiB, 82.41% compilation time)

julia> using BenchmarkTools

julia> @benchmark main()
BenchmarkTools.Trial: 323 samples with 1 evaluation.
 Range (min … max):  12.555 ms … 297.110 ms  ┊ GC (min … max):  0.00% … 92.31%
 Time  (median):     13.754 ms               ┊ GC (median):     0.00%
 Time  (mean ± σ):   15.470 ms ±  16.230 ms  ┊ GC (mean ± σ):  10.89% ± 12.55%

  ▄▆▇█▇▄                                                        
  ███████▇▄▄▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▅▅▇▄▁▅▄▁▄▁▁▁▁▁▄ ▆
  12.6 ms       Histogram: log(frequency) by time      33.9 ms <

 Memory estimate: 4.90 MiB, allocs estimate: 40064.