[CIG-SHORT] question about the convergence of Pylith 1.4.2

[Brad Aagaard]

Faqi-

If the problem converges faster with v1.4.2 than v1.3.1, you need to
adjust the scaling parameters for the nondimensionalization. This is the
only difference related to the solver between the two versions.

The default scales for the pylithapp.problem.normalizer component are

    length_scale= <dimensional>: Value to nondimensionalize length scale.
        default value: 1000*m
        current value: 1000*m, from {default}
        validator: (greater than 0*m)
    relaxation_time=<dimensional>: Relaxatim time to nondimensionalize time.
        default value: 3.15576e+07*s
        current value: 3.15576e+07*s, from {default}
        validator: (greater than 0*s)
    shear_modulus=<dimensional>: Shear modulus to nondimensionalize
pressure.
        default value: 3e+10*m**-1*kg*s**-2
        current value: 3e+10*m**-1*kg*s**-2, from {default}
        validator: (greater than 0*m**-1*kg*s**-2)

You probably need to adjust the shear_modulus or the length_scale to the
appropriate scale of your problem. Start by setting the length_scale to
1.0*m as the conditioning in v1.3.1 did not use a length scale (it only
conditioned the problem using the shear modulus).

Regards,
Brad


Faqi Diao wrote:
>   Perhaps the slower convergence is induced by the poor cell
> quality. But, I found the the computation convergents faster
> in version 1.3.1 compared with version 1.4.2, though the mesh 
> and parameter setting are the same. The only difference of the
> two model is that the Pylith 1.4.2 is of 64 bit while Pylith 1.3.1
> is of 32 bit. Maybe the aspect ratio isn't the only reason for
> the slow convergence. 
>  Please find the attached file to see the output during the iteration
>  Cheers, 
>  Faqi  
> ========================================================
>  Institute of Geodesy and Geophysics,    
>  Chinese Academy of Sciences               
>  340 Xudong St.,Wuhan, 430077, P. R. China                   
>  E-mail: faqidiao at asch.whigg.ac.cn                  
> ========================================================
>> Faqi-
>>
>> Charles may have already responded to your question (we discussed it
>> yeaterday). The mesh quality is not okay. Ideally the aspect ratio (in
>> LaGriT) should be greater than about 0.3.  Aspect ratios smaller than
>> that will degrade the convergence rate due to ill-conditioning. You
>> might be able to get away with aspect ratios of about 0.2  and still get
>> the solution to converge with a reasonable runtime.
>>
>> For users of CUBIT, the aspect ratio is the inverse of that in LaGriT,
>> so the desired range of aspect ratios is 1 to 3 (see the CUBIT
>> documentation).
>>
>> Regards,
>> Brad
>>
>>
>> Faqi Diao wrote:
>>> Hi Brad,
>>>
>>>   I found the solution converges extremly slowly in my model.  And
>>> the mesh quality seems ok (from Carl). I want to know if there are 
>>> other reasons caused the slow convergence. I list the mesh quality
>>> and petsc setting in the Pylith control file and output during the
>>> iterations. 
>>>   I use uniform time step with  'total_time = 10.0*year' and 'dt =
>>> 0.08*year'. I just want to model the postseismic deformation following
>>> an inverse earthquake with a Maxwell solid (1.0e18 Pa*s).
>>>   I think the iterations is long overdue to stop, but it didn't until
>>> the ksp_max_it (5000). Why?
>>>   Many thanks!
>>>   Best wishes, 
>>>   Faqi Diao 
>>> ========================================================
>>>  Institute of Geodesy and Geophysics,    
>>>  Chinese Academy of Sciences               
>>>  340 Xudong St.,Wuhan, 430077, P. R. China                   
>>>  E-mail: faqidiao at asch.whigg.ac.cn                  
>>> ========================================================
>>>
>>> *  mesh quality (I use LaGrit)
>>> epsilonl, epsilonaspect:   6.3155981E-07  2.5190887E-19                         
>>> --------------------------------------------                                    
>>> elements with aspect ratio < .01:                    0                          
>>> elements with aspect ratio b/w .01 and .02:         20                          
>>> elements with aspect ratio b/w .02 and .05:         98                          
>>> elements with aspect ratio b/w .05 and .1 :       2051                          
>>> elements with aspect ratio b/w .1  and .2 :      99337                          
>>> elements with aspect ratio b/w .2  and .5 :     511310                          
>>> elements with aspect ratio b/w .5  and 1. :       6806                          
>>> min aspect ratio =  0.1829E-01  max aspect ratio =  0.9561E+00                  
>>> -----------------------------------------------------------                     
>>>     619622 total elements evaluated.    
>>> *
>>> *
>>> [pylithapp.petsc]
>>> ksp_rtol = 1.0e-06
>>> pc_type = asm
>>> ksp_monitor = true
>>> ksp_view = true
>>> log_summary = true
>>> ksp_max_it = 5000
>>> ksp_gmres_restart = 50
>>> *
>>> 1728 KSP Residual norm 2.266941390044e-07 
>>> 1729 KSP Residual norm 2.265865061317e-07 
>>> 1730 KSP Residual norm 2.265604302410e-07 
>>> 1731 KSP Residual norm 2.265070386609e-07 
>>> 1732 KSP Residual norm 2.263999483430e-07 
>>> 1733 KSP Residual norm 2.263990528726e-07 
>>> 1734 KSP Residual norm 2.263126624750e-07 
>>> 1735 KSP Residual norm 2.262775558114e-07 
>>> 1736 KSP Residual norm 2.262329998822e-07 
>>> 1737 KSP Residual norm 2.261092033311e-07 
>>> 1738 KSP Residual norm 2.261048923957e-07 
>>> 1739 KSP Residual norm 2.260067875000e-07 
>>> *
>>>
>>> 　　　　　　　　　
>