DNMY: MOI

[odow]

This PR updates GLPK for the latest version of MOI, and copies the files from https://github.com/JuliaOpt/MathOptInterfaceGLPK.jl.

Blockers

• Tag a release prior to merging this
• Decide what to do with LinQuadOptInterface
• Requires RFC: various fixes for the latest MOI JuliaOpt/LinQuadOptInterface.jl#5
• Requires WIP: tidy JuliaOpt/LinQuadOptInterface.jl#6
• Requires RFC: refactor the solver-facing API JuliaOpt/LinQuadOptInterface.jl#7
• Requires support setting a single-variable objective by promoting to ScalarAffineFunction JuliaOpt/LinQuadOptInterface.jl#9
• Requires WIP: fix issues identified by MOI #354 JuliaOpt/LinQuadOptInterface.jl#10
• Requires A struct for CSRMatrix JuliaOpt/LinQuadOptInterface.jl#17
• Requires Add get_quadratic_constraint and fix getting of quadratic function JuliaOpt/LinQuadOptInterface.jl#24
• Requires Update LQOI to reflect refactoring of modifications in MOI JuliaOpt/LinQuadOptInterface.jl#29
• Requires RFC: unit tests MathOptInterface.jl#329
• Requires RFC: Comprehensive testing of basic constraint functionality MathOptInterface.jl#354
• Update MOI tag in REQUIRE

Questions

• Should we move GLPKMathProgInterface.jl to this repo?

[mlubin]

@carlobaldassi, note this is reverting our decision from 2013 to split off the JuMP-related code into a separate package. Things have changed a bit since then :).
Our current preference is to host the MathOptInterface code into the solver-wrappers for ease of maintainability and testing, unless there's a good reason not to.

@odow I don't see a need to move the code from GLPKMathProgInterface into here. It's just extra churn for code that soon won't be relevant.

[odow]

@mlubin Do you want to tag a release before this PR gets merged (or give me commit access :) )?

[mlubin]

@odow, all yours :)

[rdeits]

It looks like there's an attribute missing from the GLPKOptimizerLP(). Using this branch (and JuMP and MOI master), I can set up a simple LP:

m = Model(optimizer=GLPKOptimizerLP())
@variable m x >= 0
@objective m Min x
JuMP.optimize(m)

which fails with:

Failed to copy model into optimizer: Unsupported attribute MathOptInterface.ObjectiveFunction{MathOptInterface.SingleVariable}()

Stacktrace:
 [1] optimize!(::MathOptInterface.Utilities.CachingOptimizer{MathOptInterface.Utilities.UniversalFallback{JuMP.JuMPMOIModel{Float64}}}) at /home/rdeits/locomotion/explorations/jump-lcp/packages/v0.6/MathOptInterface/src/Utilities/cachingoptimizer.jl:155
 [2] #optimize#95(::Bool, ::Function, ::JuMP.Model) at /home/rdeits/locomotion/explorations/jump-lcp/packages/v0.6/JuMP/src/optimizerinterface.jl:48
 [3] optimize(::JuMP.Model) at /home/rdeits/locomotion/explorations/jump-lcp/packages/v0.6/JuMP/src/optimizerinterface.jl:37

If I convert the objective from a single variable to a Linear with:

@objective m Min x + 1

then everything works.

[odow]

@rdeits yes, single variable objectives aren't supported yet. They need to be implemented in LQOI.

There is a variety of other stuff not implemented yet as well. See
JuliaOpt/LinQuadOptInterface.jl#8

[rdeits]

Got it, thanks!

[coveralls]

Coverage decreased (-78.5%) to 0.0% when pulling b075666 on odow:oscar into 0da1a79 on JuliaOpt:master.

[odow]

Hmm @mlubin @joaquimg we have a problem. Because the MOI tests incorrectly allow adding multiple integer constraints (see jump-dev/MathOptInterface.jl#404, JuliaOpt/LinQuadOptInterface.jl#30, JuliaOpt/LinQuadOptInterface.jl#31), this PR now fails tests.

We can either

• disable the failing tests here
• fix MOI tests and push a new release
• merge as is with failing tests.

Thoughts?

[mlubin]

Both "disable the failing tests here" (with a TODO explaining why) and "fix MOI tests and push a new release" are viable options.

[odow]

Okay this is failing v0.7 but now passing tests on 0.6

[blegat]

We can target v0.7 for a next release

[odow]

This also broke 32-bit because of jump-dev/MathOptInterface.jl#377