[Democritus]

Hello Foamers~!

Here I met a serious problem: when i want to generate a mesh with grid number of 1e8, blockMesh just breaks down. Here is my mesh dict:

vertices
(
    (0    0     0)        //0
    ($Lx  0     0)        //1
    ($Lx  $Ly   0)        //2
    (0    $Ly   0)        //3
    (0    0     $Lz)      //4
    ($Lx  0     $Lz)      //5
    ($Lx  $Ly   $Lz)      //6
    (0    $Ly   $Lz)      //7
);

blocks
(
  hex ( 0 1 2 3 4 5 6 7 ) (10000 10000 1) simpleGrading (10 10 1)
);

Is there a way to deal with that problem? Thanks!

[akidess]

"Breaks down" is a very unspecific description. I'd blame your memory.

[jherb]

Just a complete shot in the dark: Did you compile OpenFOAM with 32 or 64 bit label size?

see e. g. http://openfoamwiki.net/index.php/Label and https://github.com/OpenFOAM/OpenFOAM...etc/bashrc#L81

[Democritus]

Originally Posted by akidess

"Breaks down" is a very unspecific description. I'd blame your memory.

Thank you very much for your attention on my problem
I now describe my situation in more details:
As you predicted, the memory ran out quickly and the hard disk was busy. Then the program do not response any more and I have to shut it down forcely. I think blockMesh generates the mesh information all in memory and do not write to disk until finished computing. So the memory become the bottleneck of my system.

Here is my idea about how to solve it:
1. generate the coarse mesh
2. use multi-thread tool in order to decompose the mesh to different sub mesh
3. refine every sub meshes in parallel

Would you please give me some advices about how to make it?
Thanks very much!

[Democritus]

Originally Posted by jherb

Just a complete shot in the dark: Did you compile OpenFOAM with 32 or 64 bit label size?

see e. g. http://openfoamwiki.net/index.php/Label and https://github.com/OpenFOAM/OpenFOAM...etc/bashrc#L81

Thank you very much for your attention on my problem
I work with ubuntu 16.04 and I installed OpenFOAM by the apt command. So I think my label set is the default value

Is there a way to install OpenFOAM with label value equaling to 64 by apt command?

Thank you very much!

[akidess]

Yes, you can go that way. Some ideas here:
http://www.cfd-online.com/Forums/ope...eneration.html

Originally Posted by Democritus

Thank you very much for your attention on my problem
I now describe my situation in more details:
As you predicted, the memory ran out quickly and the hard disk was busy. Then the program do not response any more and I have to shut it down forcely. I think blockMesh generates the mesh information all in memory and do not write to disk until finished computing. So the memory become the bottleneck of my system.

Here is my idea about how to solve it:
1. generate the coarse mesh
2. use multi-thread tool in order to decompose the mesh to different sub mesh
3. refine every sub meshes in parallel

Would you please give me some advices about how to make it?
Thanks very much!

[Democritus]

Originally Posted by akidess

Yes, you can go that way. Some ideas here:
http://www.cfd-online.com/Forums/ope...eneration.html

Dear akidess,
I have tried the 3-step-Meshing:
1. blockMesh a coarse mesh and topoSet a cellSet
2. decomposePar for making it ready for parallel processing
3. mpirun -np 8 refineMesh -overwrite -parallel
This works fine until the memory bottleneck was met
Now it comes to me a idea that if i can refineMesh one subMesh by one subMesh. I mean, if it is possible that i refine one subMesh once but still parallelly. I guess if the meshing domain could be limited to one subMesh and exploited all cpu cores' power, then the memory needed is affordable for me and the performance is ok.

[akidess]

If you run on a single node, of course running in parallel will not gain you anything in terms of memory. If you don't have more than a single node, you are out of luck I think. Yeah, you can mesh subdomains separately and then stich, but you'll probably still dump once you stitch and get the large mesh, or finally when you attempt to solve.

[Democritus]

Originally Posted by akidess

If you run on a single node, of course running in parallel will not gain you anything in terms of memory. If you don't have more than a single node, you are out of luck I think. Yeah, you can mesh subdomains separately and then stich, but you'll probably still dump once you stitch and get the large mesh, or finally when you attempt to solve.

Thank you~! I am trying to use other way like mesh grading to solve the problem.