large performance regression in calling JuMP.build using MathProgBase in 1.0

[ExpandingMan]

In my case from a few seconds to almost 2 minutes. I know it's coming from the call to JuMP.build on the model, but I don't know where yet.

Sorry for the lack of details, will look into it as soon as I get the chance, wanted to open this issue to keep track.

[ExpandingMan]

Ok, first update it seems to be slow on MathProgBase.loadproblem! which happens in JuMP solvers.jl line 372.

[odow]

That's weird because very little has changed in Gurobi.jl with respect to MathProgBase. However, there are known performance issues with LQOI using Gurobi via MOI at the moment.

[ExpandingMan]

Ok, finally traced this down and the result is a bit scary.

The regression is due to a call of sparse(transpose(A)) where A is sparse here. This regression apparently scales very badly with the size of the matrix, which in my case is (97864, 161284).

The function call in question is in SparseArrays in sparsematrix.jl somewhere in the call to SparseMatrixCSC at line 374.

At first glance I'm not at all sure about what to do as a work-around, but we'll definitely need something or this will never work for large problems in 1.0.

[ExpandingMan]

Ah, so happy there is a simple fix! This could have been quite the headache to fix. PR incoming.

[ExpandingMan]

Will be fixed by #153