Pardiso leaks memory

[termi-official]

Describe the bug 🐞

Pardiso leaks memory. This is an issue for long-running simulations with potentially thousands of linear solve calls.

Expected behavior

Pardiso should not leak memory.

Minimal Reproducible Example 👇

See this comment in the extension

LinearSolve.jl/ext/LinearSolvePardisoExt.jl

Line 150 in 6f8c1cf

# Pardiso.set_phase!(cache.cacheval, Pardiso.RELEASE_ALL)

Additional context

I am pretty sure we should call phase -1 after each solve (or at least phase 0) to release the internal buffers.

[ChrisRackauckas]

I think we need to set the finalizer? Not after each solve but on gc?

[termi-official]

For small problems that should be sufficient. However, if one solves larger linear systems with some bandwidth we potentially leak in the order of 50mb per solve call (due to the fill-in). If we use this inside a Newton for DAE solvers, then the memory consumption gets out of hand quite quickly if the time interval is not small.

Maybe a compromise is to add some flag to the constructor to decide on the release frequency?

[ChrisRackauckas]

Well we should free immediately when we can, but if say someone's REPL stays open or something, it should still have a finalizer trigger.

[j-fu]

May be this can be done on the Pardiso.jl side ? Let us see how the discussion plays out there. It seems that multiple pardiso call with RELEASE_ALL don't harm, but not sure if we can rely on this.