Efficiency of MOSEK interface

[aszekMosek]

The MOSEK direct interface inputs variables and constraints into the MOSEK task object one by one. The generic method _add_block inherited from the abstract direct solver calls the methods _add_var and _add_constraint for each variable/constraint. It is not particularly efficient, and especially in MOSEK it would be much faster to input at least the whole linear data in one go.

Here at MOSEK we wanted to see if we could achieve something like that, for instance by redefining _add_block in the MOSEK interface. So I wanted to ask if this is not against some design principle (as far as I can see all the direct interfaces follow the one-by-one pattern), if you have some suggestions how one would go about it, or if there already are any plans of this kind.

[michaelbynum]

@aszekMosek The purpose of the one-by-one pattern is for support of persistent interfaces. I think I made a poor design decision when I first wrote the direct/persistent interfaces. I should not have made them so tightly coupled. Our direct/persistent interface to CPLEX had similar issues until it was recently refactored in #1416 to cache constraints and hand them to the solver in chunks. If we ever want a persistent interface to MOSEK, then we will probably need the same sort of design. For a purely direct interface, there is no need for the one-by-one pattern at all.

I am not aware of any current plans. Furthermore, we would greatly appreciate this improvement.