Parallel model generation in JuMP




I am kinda re-iterating on The application I had in mind with when opening the issue was a large-scale binary assignment problem where most of the coefficients are integrals. Such problems arise, e.g., in statistics when you consider optimal exact tests for discrete outcomes. Essentially any expensive function for coefficient computation results in the problem, that model creation takes huge amounts of time and coefficient computation could easily be run in parallel. So, there shouldn’t be any major issues with allowing the JuMP macros to run in parallel, should there?


I would recommend doing the expensive coefficient calculations in parallel before calling JuMP’s macros. JuMP’s macros are not the bottleneck and have no need to run in parallel, so I don’t intend to consider the extra complexity of introducing (I assume) synchronization primitives into JuMP to allow them to do so.