[MITgcm-support] cg3d problem

[Nicolas Grisouard]

Dear all,

I am using a configuration, in which the code is slowly converging and 
am wondering if there is something to do about it. I tried to run a test 
with the -ftrace option activated, the resulting file is attached.

To sum up, the cg3d routine is using 75% of the time, which is maybe 
usual, I don't know. This is my first question : I am using a 2D 
configuration, in this case the size of my grid is nx*ny*nz = 
1000*1*223. Why is there a 3d process involved here? Is it the fact that 
the boundary conditions north and south are periodic? (OBCs are on)

Now, is there a way to avoid such a long convergence of the 3D solver? 
Actually, in the data file, I set these lines:

# Elliptic solver parameters
&PARM02
cg2dMaxIters=10000,
cg2dTargetResidual=1.E-13,
cg3dMaxIters=10000,
cg3dTargetResidual=1.E-13,
&

And in spite of these high values of MaxIters, the cg3d routine does not 
converge, as can be read in the STDOUT file :

(PID.TID 0000.0001)                    cg2d_init_res =    
1.40478312751614E+00
(PID.TID 0000.0001)                       cg2d_iters =  1778
(PID.TID 0000.0001)                         cg2d_res =    
9.75881835078007E-14
cg3d: Sum(rhs),rhsMax =  -1.12229402786101E-05  1.57274284756183E-12
(PID.TID 0000.0001)                    cg3d_init_res =    
4.88713802210862E+00
(PID.TID 0000.0001)                       cg3d_iters = 10000
(PID.TID 0000.0001)                         cg3d_res =    
2.38471462841502E-05


I hope there is something to do about it, instead the cost of these 
simulations will get prohibitive.

Thanks for the help,

Nicolas

-- 
Nicolas GRISOUARD
PhD Student - ERES research group
Laboratoire des Ecoulements Geophysiques et Industriels
BP 53
38041 Grenoble cedex 9 France
tel : +33 (0)476 825 037 - fax : +33 (0)476 827 022

-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: mar_tst02.o16716
URL: <http://mitgcm.org/pipermail/mitgcm-support/attachments/20081114/92c856f2/attachment.el>