Gmres fails to converge, but LU is fine

sbabbi · April 10, 2013, 21:01

Hi,
I am working on a finite-element simulation for the 3D prandtl boundary layer equations on a flat plate, using fenics.
I am using P2-P1 elements in the plane normal to the flow, and implicit euler along the stream direction.
Everything works fine if i use the LU method to factorize the Jacobian in the newton solver. On the other hand, gmres fails to converge on the very same matrix.
Now, to my understanding of gmres, there is no way that it can fail if the LU solver converge (am I wrong?).
I tested the gmres solver in other problems, and it works fine. I also tried to coarse the grid and to increase the restart parameter, with no luck.
Are there any particular reasons for which gmres may fail to converge, or is it a bug in my code?
Thanks in advance for help.

Jonas Holdeman · April 11, 2013, 09:54

I have experienced a similar problem in MATLAB and will give you my solution. My lines of code are,

% Parameters for GMRES solver
GM.Tol=1.e-13;
GM.MIter=20;
GM.MRstrt=6;
% parameters for ilu preconditioner
su.type = 'ilutp';
su.droptol = 2.e-5;
[Lm,Um] = ilu(Mat,su); % incomplete LU
Qr = gmres(Mat,RHS,GM.MIter,GM.Tol,GM.MRstrt,Lm,Um,[]); % GMRES

When the drop tolerance su.droptol is too large, GMRES does not converge. Making it sufficiently small has worked in all cases.

According to the documentation, ilutp is supposed to be the default, but the MATLAB code does not follow the documentation in ver. R2012b and must be set as above. I have not checked to see if this has been corrected in R2013a.
The parameter values are those I use in one program and vary for other programs.

April 10, 2013, 21:01	Gmres fails to converge, but LU is fine	#1
sbabbi New Member Ennio Barbaro Join Date: Apr 2013 Location: Pisa Posts: 1 Rep Power: 0	Hi, I am working on a finite-element simulation for the 3D prandtl boundary layer equations on a flat plate, using fenics. I am using P2-P1 elements in the plane normal to the flow, and implicit euler along the stream direction. Everything works fine if i use the LU method to factorize the Jacobian in the newton solver. On the other hand, gmres fails to converge on the very same matrix. Now, to my understanding of gmres, there is no way that it can fail if the LU solver converge (am I wrong?). I tested the gmres solver in other problems, and it works fine. I also tried to coarse the grid and to increase the restart parameter, with no luck. Are there any particular reasons for which gmres may fail to converge, or is it a bug in my code? Thanks in advance for help.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Interfoam blows on parallel run	danvica	OpenFOAM Running, Solving & CFD	16	December 22, 2012 03:09
Question on ILU(0) Preconditioned GMRES	Shenren_CN	Main CFD Forum	2	August 6, 2012 20:19
inviscid 2d computation for M=0.3 flow past cylinder on fine mesh can not converge	boubalos	Main CFD Forum	3	March 20, 2011 05:26
CFX FAQ updated - my simulation fails to converge	Glenn Horrocks	CFX	0	September 1, 2007 08:33
how to converge the solution for fine mesh	kathiravan	Siemens	5	August 11, 2006 02:30

April 11, 2013, 09:54		#2
Jonas Holdeman Senior Member Jonas T. Holdeman, Jr. Join Date: Mar 2009 Location: Knoxville, Tennessee Posts: 128 Rep Power: 18	I have experienced a similar problem in MATLAB and will give you my solution. My lines of code are, % Parameters for GMRES solver GM.Tol=1.e-13; GM.MIter=20; GM.MRstrt=6; % parameters for ilu preconditioner su.type = 'ilutp'; su.droptol = 2.e-5; [Lm,Um] = ilu(Mat,su); % incomplete LU Qr = gmres(Mat,RHS,GM.MIter,GM.Tol,GM.MRstrt,Lm,Um,[]); % GMRES When the drop tolerance su.droptol is too large, GMRES does not converge. Making it sufficiently small has worked in all cases. According to the documentation, ilutp is supposed to be the default, but the MATLAB code does not follow the documentation in ver. R2012b and must be set as above. I have not checked to see if this has been corrected in R2013a. The parameter values are those I use in one program and vary for other programs.