Re: [deal.II] Re: Amesos_Superludist with TrilinoWrappers bad performance

Marek, 2016-07-22 2:42 GMT-04:00 Vinetou Incucuna : > Therefore ad1) is split into small subsystems (approximately 2000 > unknowns for each core) and therefore the Amesos_Superludis should > have worked nicely? Yes SuperLU should work on this problem but I think that there is some confusion here.

Re: [deal.II] Re: Amesos_Superludist with TrilinoWrappers bad performance

On Fri, Jul 22, 2016, at 07:20 CDT, Vinetou Incucuna wrote: > Matthias, just to make things clears. What everything > should be build in release mode - deal.ii, trilinos (and > dependency libraries Blas, Lapack) > , SuperLU_Dist ? > I will give a try on Dofs renumbering. The deal.II library is

Re: [deal.II] Re: Amesos_Superludist with TrilinoWrappers bad performance

Matthias, just to make things clears. What everything should be build in release mode - deal.ii, trilinos (and dependency libraries Blas, Lapack) , SuperLU_Dist ? I will give a try on Dofs renumbering. Thank You M 2016-07-22 13:43 GMT+02:00 Jean-Paul Pelteret : > Also, have you tried optimisin

Re: [deal.II] Re: Amesos_Superludist with TrilinoWrappers bad performance

Also, have you tried optimising the bandwidth of the sparse system by renumbering the DoFs ? On Friday, July 22, 2016 at 12:30:01 PM UTC+2, Matthias Maier wrote: > > > I have the feeling, that my setup is the root cause

Re: [deal.II] Re: Amesos_Superludist with TrilinoWrappers bad performance

> I have the feeling, that my setup is the root cause of poor performance > of both solve and assemble (3s for assemble in release mode for 2000 dofs > is according to my opinion slow [17s in debug mode]). Have you run your computations against the debug library, or against the optimized release l

Re: [deal.II] Re: Amesos_Superludist with TrilinoWrappers bad performance

Hello, I have the feeling, that my setup is the root cause of poor performance of both solve and assemble (3s for assemble in release mode for 2000 dofs is according to my opinion slow [17s in debug mode]). Could You look at following stripped snippets of setup, solve and assemble, and give me ad

Re: [deal.II] Re: Amesos_Superludist with TrilinoWrappers bad performance

Hello, once more again > > 71874 dofs for Cahn-Hilliard part of system, cca 2000-3000 dofs per computer > core I have solved the Navier-Stokes-Cahn-Hilliard system on 32 cores as decoupled. Algorithm: 1)assemble, solver Cahn-Hilliard->71874 dofs 2)assemble,solve Navier-Stokes-> 2773956

Re: [deal.II] Re: Amesos_Superludist with TrilinoWrappers bad performance

Marek, 2016-07-21 15:59 GMT-04:00 Vinetou Incucuna : > I suppose, that the next > step will be an solver for general systems, > like > TrilinosWrappers::SolverBicgstab or TrilinosWrappers::SolverGMRES > ? Yes that's right. However, the performance of Krylov solvers depends on good preconditioni

Re: [deal.II] Re: Amesos_Superludist with TrilinoWrappers bad performance

Hello, thank you for the answer. I suppose, that the next step will be an solver for general systems, like TrilinosWrappers::SolverBicgstab or TrilinosWrappers::SolverGMRES ? M 2016-07-21 21:39 GMT+02:00 Bruno Turcksin : > Marek, > > On Thursday, July 21, 2016 at 3:17:48 PM UTC-4, Marek Čap

[deal.II] Re: Amesos_Superludist with TrilinoWrappers bad performance

Marek, On Thursday, July 21, 2016 at 3:17:48 PM UTC-4, Marek Čapek wrote: > > > Hello, > I have set up the Navier-Stokes-Cahn-Hilliard system in way of > step-40 with TrilinosWrappers backend. > I have assembled the phase-field part of system in reasonable time 3s, > however > the solve with Ames