Performance regression building LP in v1.14.0

[odow]

First reported by jump-dev/HiGHS.jl#337

Reproducible MPS:

medium.zip

v1.14.0

julia> using JuMP, HiGHS

julia> begin
           model = read_from_file("/Users/odow/Downloads/medium.mps")
           set_optimizer(model, HiGHS.Optimizer)
           @time MOI.Utilities.attach_optimizer(model)
       end
Running HiGHS 1.14.0 (git hash: 7df0786de3): Copyright (c) 2026 under Apache 2.0 license terms
Using BLAS: blastrampoline 
 19.168187 seconds (18.29 M allocations: 989.328 MiB, 2.97% gc time)

v1.13.1

julia> using JuMP, HiGHS

julia> begin
           model = read_from_file("/Users/odow/Downloads/medium.mps")
           set_optimizer(model, HiGHS.Optimizer)
           @time MOI.Utilities.attach_optimizer(model)
       end
Running HiGHS 1.13.1 (git hash: 1d267d97c): Copyright (c) 2026 under Apache 2.0 license terms
Using BLAS: blastrampoline 
  1.895229 seconds (18.29 M allocations: 989.327 MiB, 33.30% gc time)

Flamegraph

Flamegraph of v1.14 shows time is dominated by calls to Highs_addRow:

[odow]

With PProf I get this trace:

where previously it was

[odow]

Which seems to point to #2857

and
https://github.com/ERGO-Code/HiGHS/pull/2857/changes#diff-464243dd6f328f3da7dc75290f811765e6fab1a862510ec11a0053020fb7bdb2R150

HiGHS/highs/util/HighsMatrixUtils.cpp

Line 150 in dcc2530

el_in_vec.assign(vec_dim, illegal_el);

[jajhall]

Indeed, when creating a model by adding each constraint as a single row this will be expensive. It stems from the new need to add Hessian duplicates rather than just detecting them and returning an error. I can fix it.

I'm puzzled by it being reproducible with an MPS file. When JuMP has read an MPS file, surely it passes the constraint matrix to a solver in a single call.

[odow]

It can, but I haven’t implemented the corresponding method in HiGHS.jl so it falls back to the row by row.

[jajhall]

Closed by #2984