MOOSE/Moltres memory usage

[sychang2021]

Hi, everyone.

I want to calculate a case with a lot of variables. It has 46,800 elements, 17 variables and 570,000 degrees of freedom.

I used the lu method and the Executioner settings are as follows:

  type = Transient
  num_steps = 100
  solve_type = 'NEWTON'
  dt = 2e-7
  #end_time = 100

  nl_rel_tol = 1e-6
  nl_abs_tol = 1e-6

  solve_type = 'NEWTON'
  petsc_options = '-snes_converged_reason -ksp_converged_reason -snes_linesearch_monitor'
  petsc_options_iname = '-pc_type -pc_factor_shift_type'
  petsc_options_value = 'lu       NONZERO'
  line_search = 'none'
   # petsc_options_iname = '-snes_type'
  # petsc_options_value = 'test'

  nl_max_its = 30
  l_max_its = 100
  #steady_state_detection = true
  #steady_state_tolerance = 1e-10

  automatic_scaling = true

In fact this is a small single rod model and the number of elements is very small. Even so， when I run it on a computer with 64GB of RAM, the RAM is full and no results are produced. If I want to use MOOSE to calculate relatively large problems, it is obviously not feasible, such as the problem of six groups of neutrons and six groups of recursors with millions of grids.

I have tried ilu but the results do not converge, and only when using lu (maybe set to other options will also converge). Can you give me some advice on saving memory and guaranteeing convergence?

[sychang2021]

There is another question about scaling . I want to use automatic_scaling = true in Executioner block. At the beginning of the neutron calculation, if the neutron flux and the precursors have small values, their residuals are very small. Will this cause MOOSE to think that the iteration has converged, which will affect the accuracy of the solution?

[smpark7]

64GB RAM shared by 8 processors should be more than enough for a problem with 570,000 degrees of freedom. You may refer to this page for recommendations on the petsc options. asm with lu as the subsolver will require less memory. ilu is usually not good enough to resolve the strong coupling between the flux and precursor variables.

automatic_scaling is good for the problem you're describing. If the flux and precursor values are small, automatic_scaling scales their residual or jacobian values up to approximately unity (~1). You'll also want compute_scaling_once = false so that Moltres adjusts the scaling factor at every timestep.