Performance tips

There are three phases of building and solving a Convex.jl problem:

  1. Building the expression tree. This is everything that happens before solve! is called, for example, creating variables, problems, constraints, and other expressions.
  2. Formulating the problem in MathOptInterface. This happens automatically when solve! is called, and (as long as solve! is not called with silent=true) emits a log message when completed, containing the amount of time it took and the total amount of memory allocations.
  3. Solving the problem. This happens in solve! after the problem is formulated. When silent=false, typically solvers will print information about the problem and their progress when solving it.

If you are experiencing performance issues, it is important to isolate which of these three steps is taking the most time. Typically, step (1) should be the fastest, then step (2), then step (3).

If you're struggling to improve performance, feel free to ask for help in the community forum. Before asking a question, make sure to read the post make it easier to help you, which contains a number of tips on how to ask a good question.

Faster problem formulation: Avoid scalar indexing

When manipulating Convex.jl expressions, try to use a "vectorized" style, and avoid loops and scalar indexing. For example:

function bad_my_dot(x, y) # avoid this kind of code for Convex.jl expressions!
    s = 0.0
    for i in eachindex(x, y)
        s += x[i]*y[i]
    return s

is usually perfectly reasonable Julia code (although perhaps s should be initialized as s=zero(promote_type(eltype(x), eltype(y)))) to be more generic), since loops are typically fast in Julia.

However, when using Convex.jl expressions like x = Variable(100), calling bad_my_dot(x, y) with y = rand(100) will create 100 separate IndexAtoms, each of which has some overhead. That is because Convex maintains a lazy expression tree representing the problem. Instead, the vectorized form sum(x .* y) or, y) will be more efficient.

Indexing style typically should have no effect on solve times (step (3) above), only on constructing the expression tree and formulating the problem.

Speeding up solve times: dualization

Depending on the solver and the problem, it can sometimes speed things up quite a lot to pass the solver the dual problem to solve rather than the primal problem. This can be accomplished easily with Dualization.jl, simply by passing dual_optimizer(optimizer) instead of optimizer to solve!. See Dualization for an example.

Speeding up solve times: digging into the final formulation

Convex.jl is built around conic programming, and it works with conic solvers. This may be a different optimization methodology than you may be used to, and it works by formulating problems in terms of affine objective functions and affine-function-in-cone constraints for a variety of convex cones. Convex.jl reformulates all problems to this form. For example, the objective function you give Convex will not actually be executed line-by-line; instead, it will be reformulated to a possibly very-different looking form for the conic solver to handle.

Convex.jl provides an abstraction by automatically performing these reformulations, but as often is the case, performance leaks through. To have the best performance, you might need to learn more details about the various cones supported by your solver, by MathOptInterface, and how Convex.jl reformulates your problem. There could be mathematically equivalent formulations that are faster, and in some cases it could be that Convex's automatic reformulations can be improved.

One option to view the final problem after these reformulations is to write it to disk with Convex.write_to_file and inspect it manually or with other software.