[Trilinos-Users] ilut setting

[Erik Boman]

Unfortunately, the incomplete factorization parameters are not well 
documented so looking at the source code is best. Yes, Ifpack_ILUT 
interprets "ilut level-of-fill" to be the ratio of nonzeros in the 
preconditioner relative to the original matrix, so a value of 1.0 or a 
bit higher is reasonable. (Note: 1.0 in Ifpack roughly corresponds to 0 
in AztecOO.) If you set it to 50, you will in most cases compute a 
complete factorization (but in a very slow way!). As a sanity check, you 
can try using Amesos (by default, KLU) as your subdomain solver.

There are several other parameters in Ifpack that may be helpful. For 
example, the absolute and relative threshold parameters may make your 
preconditioner more robust. Confusingly, these prescribe a diagonal 
perturbation (aka Manteuffel shift).

Erik

Hoang Giang Bui wrote:
> Hi Chris
>
> I didn't see the range of value of ilut level of fill in Ifpack user 
> guide yet. But looking in the code of Ifpack_ILUT, I'm quite sure that 
> the level of fill indicates the number of nonzeros in the 
> preconditioner compare to initial nonzeros. Therefore its value can be 
> bigger than 1.
>
> Giang Bui
>
>
> On 09/18/2013 04:43 AM, Chris Jackson wrote:
>>
>> I don’t know if this applies to the one you are talking about, but I 
>> recall that with one of the ILUT preconditioners the fill value is 
>> expected to be between 0.0 and 1.0.  1.0 meant essentially we were 
>> finding the inverse of the matrix.  Before we figured that out we 
>> were always astounded that BiCGStab was converging in 2 iterations…
>>
>>  
>>
>> Chris Jackson
>>
>>  
>>
>>  
>>
>> *From:* trilinos-users-bounces at software.sandia.gov 
>> [mailto:trilinos-users-bounces at software.sandia.gov] *On Behalf Of 
>> *Hoang Giang Bui
>> *Sent:* Tuesday, September 17, 2013 6:28 PM
>> *To:* Erik Boman
>> *Cc:* trilinos-users at software.sandia.gov
>> *Subject:* Re: [Trilinos-Users] ilut setting
>>
>>  
>>
>> Hi Erik
>>
>> I tried to increase "fact: ilut level-of-fill" to 50.0 and it still 
>> doesn't converge. In fact, with this value the preconditioner take 
>> very long to compute. Is there other option we can set for Ifpack ILUT?
>>
>> Giang Bui
>>
>>  
>>
>> On Tue, Sep 17, 2013 at 8:01 PM, Erik Boman <egboman at sandia.gov 
>> <mailto:egboman at sandia.gov>> wrote:
>>
>> Giang Bui,
>>
>> You will not get exactly the same answer because the ILUT 
>> implementations are different in AztecOO and Ifpack. Also, the 
>> meaning of "fill" may be slightly different. Try increase the value 
>> of  "fact: ilut level-of-fill" in the Ifpack parameter to see if you 
>> can make it converge.
>>
>> Erik
>>
>> Hoang Giang Bui wrote:
>>
>> Dear Trilinos developers,
>>
>> Is there difference between 2 settings:
>>
>> solver_parameters.set("AZ_solver", "AZ_bicgstab")
>> solver_parameters.set("AZ_kspace", 1000)
>> solver_parameters.set("AZ_output", 100)
>> solver_parameters.set("AZ_precond", "AZ_dom_decomp")
>> solver_parameters.set("AZ_subdomain_solve", "AZ_ilut")
>> solver_parameters.set("AZ_drop", 0.0)
>> solver_parameters.set("AZ_ilut_fill", 10.0)
>> aztec_solver.SetParameters(solver_parameter, true);
>>
>> And:
>> solver_parameters.set("AZ_solver", "AZ_bicgstab")
>> solver_parameters.set("AZ_kspace", 1000)
>> solver_parameters.set("AZ_output", 100)
>> solver_parameters.set("AZ_precond", "AZ_none")
>> aztec_solver.SetParameters(solver_parameter, true);
>> IFPreconditionerType = "ILUT"
>> preconditioner_parameters.set("fact: ilut level-of-fill", 10.0);
>> preconditioner_parameters.set("fact: drop tolerance", 0.0);
>> overlap_level = 1
>> mPrec = Teuchos::rcp(Factory.Create(IFPreconditionerType, &rA, 
>> overlap_level));
>> mPrec->SetParameters(preconditioner_parameters);
>> aztec_solver.SetPrecOperator(&*(mpPrec->GetOperator()))
>>
>> I thought it all resulted in Additive-Schwartz preconditioner with 
>> ILUT as local subdomain preconditioner. But it gives different 
>> performance. The first setting works well but the second failed to 
>> converge.
>>
>> Best regards
>> Giang Bui
>>
>> _______________________________________________
>> Trilinos-Users mailing list
>> Trilinos-Users at software.sandia.gov 
>> <mailto:Trilinos-Users at software.sandia.gov>
>> http://software.sandia.gov/mailman/listinfo/trilinos-users
>>  
>>
>>  
>>
>>
>>
>>
>> -- 
>> With Best Regards !
>> Giang Bui
>>
>> To learn and to excel
>>
>