Performance regression from Plasmo 0.5.4 -> 0.6.4

[etatara]

I have a gas pipeline network model that has been working fine with Plasmo 0.5.4 and after updating to Plasmo 0.6.4, the number of Ipopt iterations and solve time is more than 2x longer for the same model. The only code difference in the model is that I use Plasmo.set_to_node_objectives(graph) when using Plasmo 0.6.4. The objective value and model variable solutions are 100% exact between the two versions of Plasmo. For small problems the performance difference isn't really noticeable, but on larger models with > 1E5 variables, the increase in solution time gets large.

My current environment is on linux with:

• Julia 1.10.9
• Ipopt.jl 1.8.0
• Ipopt binary 3.14.17 with ma97+MKL
• JuMP 1.25.0

It seems that Ipopt is getting the same model information to produce the same solution, but I wonder if there is some difference in the way that Plasmo is handling this differently through the graph backend between versions. I've been trying to examine the IpoptProblem (from Ipopt.jl) data but can't really come up with a way to see exactly what is different between Plasmo versions.

[jalving]

Hey @etatara, I also assume the start values are the same in both models? (both primal and dual)

It is possible that the ordering of variables and constraints is different in the Ipopt model. This can certainly happen if you're using subgraphs. If you are using subgraphs, perhaps we could try building two small equivalent problems to check the ordering. I will have a deeper look at this when I get a chance.