Changes between Version 29 and Version 30 of ControllingRefinement


Ignore:
Timestamp:
07/23/12 12:41:20 (12 years ago)
Author:
Jonathan
Comment:

—

Legend:

Unmodified
Added
Removed
Modified

  • ControllingRefinement

    v29 v30  
    22= Controlling Refinement in AstroBEAR 2.0 =
    33== Overview ==
    4 Refinement is triggered by various sections of code evaluating various refinement criteria described below.  To mark cell {{{i,j,k}}} for refinement, each routine sets {{{Info%ErrFlag(i,j,k)=1}}}.  If a cell is marked for refinement by more then one routine there is no problem since it will still be refined.  Setting {{{Info%ErrFlag(i,j,k)=0}}} however, is generally not a good idea since another flagging routine may have previously marked the cell for refinement with good reason.
    54
    6 After all of the cells in a given parent info have been potentially marked for refinement, the subroutine NewSubGrids in [source:'Scrambler (Development)/data/data_info_ops.f90' data_info_ops.f90] groups the flagged cells into rectangular patches so that every cell is within a patch and that no patch has a filling ratio less then the !DesiredFillRatio for that level set in [GlobalDataExplained global.data].  Having a !DesiredFillRatio of 1.0 will effectively limit refinement to only cells marked for refinement however the !MinimumGridSize also set in [GlobalDataExplained global.data] may limit the actual desired fill ratios.  Having many small grids is generally not as efficient as having a few larger grids even if they comprise more cells and so a desired fill ratio of 1.0 does not usually give the best performance.  Values of .7 to .9 are recommended for optimal performance.
     5Refinement of grids is controlled by RefinementObjects or by setting {{{Info%ErrFlag(i,j,k)=1}}} in the Problem modules subroutine {{{ProblemSetErrFlag(Info)}}}.  If a cell is marked for refinement by more then one routine there is no problem since it will still be refined.  Setting {{{Info%ErrFlag(i,j,k)=0}}} however, is generally not a good idea since another flagging routine may have previously marked the cell for refinement with good reason.
    76
    8 ||  Fill Ratio = .7  ||  Fill Ratio = .9  ||  Fill Ratio = 1.0  ||
    9 || [[Image(ErrFlag1_7.png, width=513)]] || [[Image(ErrFlag1_9.png, width=513)]] || [[Image(ErrFlag1_10.png, width=513)]] ||
     7After all of the cells in a given parent info have been potentially marked for refinement, the subroutine NewSubGrids in [source:'Scrambler (Development)/data/data_info_ops.f90' data_info_ops.f90] groups the flagged cells into rectangular patches so that every cell is within a patch. 
    108
    11 Additionally one can increase the minimum grid size to reduce the prevalence of long slender grids.  Below we increased the minimum refined patch to be 4 parent cells.  (This results in actual child grids that are at least 8x8)
     9* The previous version of the code (which is also used if lUseOriginalNewSubGrids = .true. in global.data) then split and shrunk grids so that no patch had a filling ratio less then the !DesiredFillRatio for that level set in [GlobalDataExplained global.data].  Having a !DesiredFillRatio of 1.0 will effectively limit refinement to only cells marked for refinement however the !MinimumGridSize also set in [GlobalDataExplained global.data] may limit the actual desired fill ratios.  Having many small grids is generally not as efficient as having a few larger grids even if they comprise more cells and so a desired fill ratio of 1.0 does not usually give the best performance.  Values of .7 to .9 are recommended for optimal performance.
    1210
    13 || [[Image(ErrFlag1_10_4.png, width=513)]]
     11 ||  Fill Ratio = .7  ||  Fill Ratio = .9  ||  Fill Ratio = 1.0  ||
     12 || [[Image(ErrFlag1_7.png, width=513)]] || [[Image(ErrFlag1_9.png, width=513)]] || [[Image(ErrFlag1_10.png, width=513)]] ||
     13
     14 Additionally one can increase the minimum grid size to reduce the prevalence of long slender grids.  Below we increased the minimum refined patch to be 4 parent cells.  (This results in actual child grids that are at least 8x8)
     15
     16 || [[Image(ErrFlag1_10_4.png, width=513)]]
     17
     18
     19* The newer default version (lUseOriginalNewSubGrids = .false.) uses an improved algorithm that takes into consideration the additional cost of smaller grids.  It splits and shrinks but stops when doing so will create more overall work.  The !DesiredFillRatio and !MinimumGridSize are ignored.
     20
     21
    1422
    1523After the new patches are determined, the distribution algorithm may split these patches to accommodate load balancing.  For example running on 3 processors we get
    … …  
    7684
    7785
    78 
    79 
    80 
    81 
    82 
    83 
    84 
    85 
    86 
    87 
    88 
    89 
    90 
    91 
    92 
    93 
    94 
    95 
    96 
    97 
    98 
    99 
    100 
    101 
    102 
    103 
    104 
    105 
    106 
    107 
    108 
    109 
    110 
    111 
    112 == Field Gradient Refinement ==
    113 AstroBEAR 2.0 comes with the ability to refine on field gradients.  For each fluid variable, AstroBEAR checks if...
    114 
    115 [[latex($\frac{|q_{i+1}-q_{i}|}{\max \left({\frac{|q_{i+1}|+|q_{i}|}{2},\mbox{MinScale}} \right)} > \mbox{tol}$)]]
    116 
    117 and marks cell {{{i+1}}} and cell {{{i}}} for refinement.  This is essentially [[latex($\Delta x \left | \frac{\partial log(|q|)}{\partial x} \right |$)]].  The [[latex($\Delta x$)]] is for the current grid resolution and results in coarser levels being more sensitive to fluid gradients for a given tolerance.  The values of tol is found by multiplying the qTolerance for the level with the refineVariableFactor for the fluid variable.  !MinScale is dependent on which field you are considering and is given in the following table:
    118 || Field || !MinScale ||
    119 || rho || 0.0 ||
    120 || Pressure || 0.0 ||
    121 || Tracer fields || [[latex($\rho$)]] ||
    122 || Bx, By, Bz || [[latex($\sqrt{\rho}c_s$)]]
    123 || Px, Py, Pz || [[latex($\rho c_s$)]]
    124 
    125 Note that strong adiabatic shocks will typically be resolved by a few [[latex($n$)]] cells depending on the method (PPM, PLM, etc...) and will be refined for density when [[latex($\mbox{tol} < \frac{(4-1)/n}{\frac{4+1}{2}} = \frac{3}{2.5n}= \frac{1.2}{n}$)]].  For PPM or PLM this requires tolerances < .3 or so to resolve grid aligned adiabatic shocks based on density gradients.  To resolve non-grid aligned shocks tolerances closer to .1 should be used.  ALso note that the pressure jumps will typically be much larger (of order [[latex($M^2$)]]) so pressure gradients will trigger refinement when [[latex($M > \sqrt{n\mbox{tol}}$)]]. 
    126 
    127 Also note that the momentum refinement is not Galilean invariant.  But then again, neither is the numerical diffusion...
    128 
    129 
    13086=== Changing qTolerance and !DesiredFillRatios ===
    13187The most direct and quickest way to change refinement is to change the {{{qTolerance}}} and !DesiredFillRatios arrays which are both located in the {{{global.data}}} file.  Each element of these arrays is specific to that level. For example, the first element is for the root level, the next element is for the 1st level of refinement, and so on. {{{qTolerance}}} effectively sets how large a gradient must be in order for those two cells to be refined.  Higher {{{qTolerance}}} --> less refinement.  !DesiredFillRatios affects the way astrobear lays down grids or patches for refinement.  For example, if the !DesiredFillRatio is 0.7, then the AMR will lay down patches so that at least 70% of each patch contains cells marked for refinement.  As a result of this process, you will often get cells being refined which were not initially marked for refinement.  In general, lower !DesiredFillRatios --> more refinement.  Note that changing these arrays is very problem specific.