View Issue Details
| ID | Project | Category | View Status | Date Submitted | Last Update |
|---|---|---|---|---|---|
| 0001542 | OpenFOAM | Bug | public | 2015-02-17 21:31 | 2015-04-09 12:54 |
| Reporter | feymark | Assigned To | henry | ||
| Priority | high | Severity | major | Reproducibility | always |
| Status | closed | Resolution | no change required | ||
| Platform | Linux | OS | Arch Linux | ||
| Summary | 0001542: Restart introduce perturbations when increasing number of processors | ||||
| Description | When running the channel395 on 8 processors restart from a previous time works perfectly fine and exactly the same residuals are found for all following time steps. However when the number of processors are increased to 16 the restart introduces perturbations and different residuals, and thus p and U, are found. | ||||
| Steps To Reproduce | Run the attached cases using Allrun. | ||||
| Tags | No tags attached. | ||||
|
|
|
|
|
Yes, this is to be expected. In order to reduce this perturbation you would need to run with tighter tolerances, and even then machine round-off may limit the degree to which you can mitigate this perturbation. The best approach is not to change the decomposition mid-run. |
|
|
I notice from you answer that you haven't looked at my files. I do not change the decomposition mid-run. I'm not a newbie. |
|
|
Are the statistics when avereged long enough to converge different between the two runs? |
|
|
Perhaps I was a bit unspecific. If you look at the final pressure iterations for the 8 processor case, starting from 0. At time = 3 it reads, GAMG: Solving for p, Initial residual = 0.75319219919945, Final residual = 3.5915425979914e-07, No Iterations 12 Then restart from time = 2. At time = 3 it reads, GAMG: Solving for p, Initial residual = 0.75319219919945, Final residual = 3.5915425979914e-07, No Iterations 12 Exactly the same, which is very good. But if you do the same thing with 16 processors you get, GAMG: Solving for p, Initial residual = 0.77177480795539, Final residual = 8.4216285201411e-07, No Iterations 11 and after restart from time = 2 GAMG: Solving for p, Initial residual = 0.77177480795462, Final residual = 8.4216285156141e-07, No Iterations 11 They are close but they are not the same. I do not mean that 8 processors should give exactly the same answer as 16 processors. But, the restart should be exact regardless of the number of processors. |
|
|
I found out what's causing this. The problem is that the uniform/time isn't written in binary although the format clearly states binary. /*--------------------------------*- C++ -*----------------------------------*\ | ========= | | | \\ / F ield | OpenFOAM: The Open Source CFD Toolbox | | \\ / O peration | Version: 2.3.x | | \\ / A nd | Web: www.OpenFOAM.org | | \\/ M anipulation | | \*---------------------------------------------------------------------------*/ FoamFile { version 2.0; format binary; class dictionary; location "2/uniform"; object time; } // * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * // value 1.9999999999999998; name "2"; index 10; deltaT 0.20000000000000001; deltaT0 0.20000000000000001; // ************************************************************************* // |
|
|
The uniform/time file is written in the format you specified in your controlDict and if you write field data into this file it would be binary, however individual scalars are written ASCII as they are in field files. I could change the uniform/time file to be explicitly written ASCII but then all data including field data would be written ASCII. |
|
|
Greetings to all! I believe this issue was addressed or at least talked about in issue #815: http://www.openfoam.org/mantisbt/view.php?id=815 - the changes are present in OpenFOAM-dev, but I think they aren't in 2.3.x, because the changes were substantial. @Henry: And if I remember correctly, the "uniform/time" file is always written in ASCII. @Feymark: Why do you need the time precision to be set to 14? Because writing floating point numbers in ASCII is always prone to error. Best regards, Bruno |
|
|
The value, deltaT and deltaT0 needs to be written in binary (or in most cases ASCII with write precision 17) for the restart to work perfectly. |
|
|
Writing these entries in binary will not improve the accuracy of the internal representation. If 17 decimal places is not sufficiently accurate for your simulations you will need to recompile OpenFOAM quad-precision or use an exact kernel. Given that you are not changing the time-step in your simulations you do not need the uniform/time file anyway and you could delete it. |
| Date Modified | Username | Field | Change |
|---|---|---|---|
| 2015-02-17 21:31 | feymark | New Issue | |
| 2015-02-17 21:31 | feymark | File Added: channel395.tar.gz | |
| 2015-02-17 21:48 | henry | Note Added: 0003812 | |
| 2015-02-17 21:48 | henry | Status | new => closed |
| 2015-02-17 21:48 | henry | Assigned To | => henry |
| 2015-02-17 21:48 | henry | Resolution | open => fixed |
| 2015-02-17 21:50 | feymark | Note Added: 0003813 | |
| 2015-02-17 21:50 | feymark | Status | closed => feedback |
| 2015-02-17 21:50 | feymark | Resolution | fixed => reopened |
| 2015-02-17 21:50 | feymark | Note Edited: 0003813 | |
| 2015-02-17 22:36 | henry | Note Added: 0003814 | |
| 2015-02-17 22:51 | feymark | Note Added: 0003815 | |
| 2015-02-17 22:51 | feymark | Status | feedback => assigned |
| 2015-04-09 10:06 | feymark | Note Added: 0004583 | |
| 2015-04-09 10:17 | henry | Note Added: 0004584 | |
| 2015-04-09 10:45 | wyldckat | Note Added: 0004586 | |
| 2015-04-09 10:46 | feymark | Note Added: 0004587 | |
| 2015-04-09 12:53 | henry | Note Added: 0004588 | |
| 2015-04-09 12:53 | henry | Status | assigned => closed |
| 2015-04-09 12:54 | henry | Resolution | reopened => no change required |
