MLMG solver fails if particles initialization is `NUniformPerCell`, not if `NRandomPerCell`

[aeriforme]

Description

In a simple 3D simulation of a uniform plasma, the electrostatic MLMG solver doesn't converge when switching particle initialization from NRandomPerCell to NUniformPerCell.
The error is: MLMG: Bottom solve failed.

How to reproduce

• WarpX inputs files: inputs_mlmg_issue.txt

System information

• Perlmutter, GPU partition, on either 1 or 4 GPUs.
• WarpX version: 25.10-113-g607b091a8295

Additional information

• Terminal output when converging (i.e. using NRandomPerCell): output_success.txt
• Terminal output when failing (i.e. using NUniformPerCell): output_fail.txt

CC: @RemiLehe

[WeiqunZhang]

If I add this, if runs fine.

--- a/Source/ablastr/fields/PoissonSolver.H
+++ b/Source/ablastr/fields/PoissonSolver.H
@@ -405,6 +405,8 @@ computePhi (
         mlmg.setMaxIter(max_iters);
         mlmg.setConvergenceNormType(amrex::MLMGNormType::greater);
 
+        rho[lev]->OverrideSync(geom[lev].periodicity());
+
         const int ng = int(grid_type == utils::enums::GridType::Collocated); // ghost cells
         if (ng) {
             // In this case, computeE needs to use ghost nodes data. So we

That suggests that the data in rho are not consistent on shared nodes. We need to understand why.

[WeiqunZhang]

It seems that OverridSync is never called after the charge is deposited in ComputeSpaceChargeField.

I am not sure whether we should call OverrideSync in computePhi or in LabFrameExplicitES::ComputeSpaceChargeField. The former is safer, but other callers might have consistent data already.

In principle, this bug is NOT limited to periodic boundary or NUniformPerCell. But when amrex solver is able to converge, it will always provide consistent phi. So if rho is not used in other places, the results of NRandomPerCell should be fine.

@RemiLehe

[ax3l]

Proposed fix in #6438