Good idea, man, but I don't think changing preconditioner can control the residual error.
2012/8/29 Vikram Garg <[email protected]>: > Yuanwu, > To set the lu preconditioner type just add the following to > the command you use to execute: -pc_type lu . If you want to use some other > kind of preconditioner you can just change lu to whatever type you want > (ilu, jacobi etc) I believe the default is jacobi. > > Warning: If your problem sizes are really large then lu might take a while. > Just try it out first and then you can try ilu if its too slow. > > Thanks. > > > On Tue, Aug 28, 2012 at 9:02 PM, 蔡园武 <[email protected]> wrote: >> >> 2012/8/29 Vikram Garg <[email protected]>: >> > Hey Yuanwu, >> > Can you tell us what kind of linear solver options >> > you >> > are using (the method, preconditioner options passed at runtime) ? >> >> I used the default option. (I don't know what's the default?) >> #EquationSystems >> n_systems()=1 >> System #0, "homogenization_elastic_3D_plate" >> Type "LinearImplicit" >> Variables="u" "v" "w" >> Finite Element Types="LAGRANGE" "LAGRANGE" "LAGRANGE" >> Approximation Orders="FIRST" "FIRST" "FIRST" >> n_dofs()=648 >> n_local_dofs()=648 >> n_constrained_dofs()=198 >> n_local_constrained_dofs()=198 >> n_vectors()=13 >> n_matrices()=1 >> DofMap Sparsity >> Average On-Processor Bandwidth <= 84 >> Average Off-Processor Bandwidth <= 0 >> Maximum On-Processor Bandwidth <= 144 >> Maximum Off-Processor Bandwidth <= 0 >> DofMap Constraints >> Number of DoF Constraints = 198 >> Average DoF Constraint Length= 1 >> >> > If you >> > are using the same stiffness matrix (K) and just changing the rhs (F), >> >> That's just what I'm doing. >> >> > then >> > you might actually be better of just using the LU preconditioner for one >> > load and reusing it for the others. >> >> How to set it in my code? Can you give me a simple example? >> Thanks for your answer! >> >> > >> > Thanks. >> > >> > On Tue, Aug 28, 2012 at 8:45 PM, 蔡园武 <[email protected]> wrote: >> >> >> >> Hi, guys, >> >> I have a LinearImplicitSystem, solved with a sequential of different >> >> force vectors (rhs). >> >> Actually I defined different 'Fe' assemble function using a 'loadcase' >> >> indicator. In main function, I set the 'loadcase' value, call >> >> system.solve(), then es.reinit(), set a new 'loadcase' value, and >> >> solve() again. >> >> But I found that for some 'loadcase', the solver converged badly, like: >> >> >> >> loadcase1: Linear solver converged at step: 10736, final residual: >> >> 2.00331e-21 >> >> loadcase2: Linear solver converged at step: 8549, final residual: >> >> 2.10685e-21 >> >> loadcase3: Linear solver converged at step: 8, final residual: >> >> 1.38269e-07 >> >> loadcase4: Linear solver converged at step: 0, final residual: 0.463112 >> >> loadcase5: Linear solver converged at step: 0, final residual: 0.463112 >> >> loadcase6: Linear solver converged at step: 0, final residual: 0.463112 >> >> >> >> In loadcase3, Linear solver converged at step 8, final resudual is not >> >> small enough. It's very strange that the loadcases after this didn't >> >> run? (converged at step 0?) The results are wrong and unbelievable. >> >> But if I solve loadcase3 after all the other loadcases, then loadcase >> >> 4,5,6 will be all right. Don't know why? >> >> >> >> Can I mannually control the final resudual tolerance? I did set a >> >> parameter in es: >> >> es.parameters.set<unsigned int> ("linear solver maximum >> >> iterations") = 20000; >> >> es.parameters.set<Real> ("linear solver tolerance") = TOLERANCE; >> >> >> >> Thanks for your help！ >> >> -- >> >> Cai Yuanwu 蔡园武 >> >> Dept. of Engineering Mechanics, >> >> Dalian University of Technology, >> >> Dalian 116024, China >> >> >> >> >> >> >> >> ------------------------------------------------------------------------------ >> >> Live Security Virtual Conference >> >> Exclusive live event will cover all the ways today's security and >> >> threat landscape has changed and how IT managers can respond. >> >> Discussions >> >> will include endpoint security, mobile security and the latest in >> >> malware >> >> threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/ >> >> _______________________________________________ >> >> Libmesh-users mailing list >> >> [email protected] >> >> https://lists.sourceforge.net/lists/listinfo/libmesh-users >> > >> > >> > >> > >> > -- >> > Vikram Garg >> > PhD Candidate >> > Institute for Computational and Engineering Sciences >> > The University of Texas at Austin >> > >> > http://users.ices.utexas.edu/~vikram/ >> > http://www.runforindia.org/runners/vikramg >> > >> >> >> >> -- >> Cai Yuanwu 蔡园武 >> Dept. of Engineering Mechanics, >> Dalian University of Technology, >> Dalian 116024, China > > > > > -- > Vikram Garg > PhD Candidate > Institute for Computational and Engineering Sciences > The University of Texas at Austin > > http://users.ices.utexas.edu/~vikram/ > http://www.runforindia.org/runners/vikramg > -- Cai Yuanwu 蔡园武 Dept. of Engineering Mechanics, Dalian University of Technology, Dalian 116024, China ------------------------------------------------------------------------------ Live Security Virtual Conference Exclusive live event will cover all the ways today's security and threat landscape has changed and how IT managers can respond. Discussions will include endpoint security, mobile security and the latest in malware threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/ _______________________________________________ Libmesh-users mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/libmesh-users
