Efficiency of the Garbage Collection

[JordiManyer]

Hello all,

@amartinhuertas and me have been working on a Geometric MultiGrid solver/preconditioner for Gridap. To solve the coarsest-level linear system in parallel, we decided to use MUMPS through a PETScLinearSolver from GridapPETSc.

While profiling the resulting algorithm for a Poisson problem, where we use a preconditioned Conjugate Gradient solver with our GMG preconditioner, we find that a weirdly high amount of time is spent in the method gridap_petsc_gc(), which is called from within solve!() at every iteration of the conjugate gradient.

Due to this, we think it could be advantageous to introduce a new flag/kwarg/setting that deactivates all these automatic calls to the GC that are scattered around the code. The idea is to give an advanced user/developer the ability to manually call the GC optimally depending on the application.

Thanks!

[amartinhuertas]

Thanks for reporting @JordiManyer ! Could you confirm if removing the call to gridap_petsc_gc() the hot spot/bottleneck disappears/is atenuated?