Home › Forums › Questions

Remeshing, Amazon EC2, Convergence Criteria, Export

Submitted by Magrn on January 25, 2011 - 23:27

1. While meshing, Caedium creates the full mesh and then goes on to remesh at a significantly lesser speed of around 200tets/sec. This happens with a mesh of 1m million tets while running on a amazon EC2 cluster (26 core, 68GB ram) but not in my colleague's laptop. (same caedium settings) [both trial versions]

2. Do you have any recommendations on how to optimise Caedium to run on a Amazon EC2 Windows cluster of 26 EC units (8cores) and 68GB (using 100% CPU, but only 6GB RAM for simulations). More specifically : should I use parallel config as 'shared memory' or 'cluster' . If so, are processes/cores 8 or 26 ?.

I've run a simulation of 0.3 million cells in 20 mins (no issues with remeshing either), but same geomtery of 0.6million cell took 6 hours on the same machine. Haven't run the simulation for 1million cells as yet due to remeshing issues.

Also, I'm planning to buy a 1year subscription to Caedium. What license(/s) would you suggest if I'm to run Caedium on a Amazon EC2 instance (26 cores, 68GB ram ).

3. Where can I change convergence criteria ?

4. I'm unable to save/export if I interrupt/pause the simulations before hitting the max number of iterations specified. The message received is something like ' Interrupted: unable to save file ............. .case '.

5. How can I normalize lift and drag values to output Cl and Cd.

Sorry for any ignorance and hope I've expressed my queries clearly.

Thanks

‹ Lift and drag Momentum Drops ›

Good Questions - Some Answers

Submitted by symscape on January 26, 2011 - 10:54.

In answer to your good questions:

  1. Sorry I have no experience running Caedium on Amazon EC2. At a guess it's likely there is a problem with the mesh somewhere (unless your colleague generated the exact same mesh) - maybe the mesh sizing is actually too small causing elements to collapse. I suggest you construct your geometry with the small elements close to the origin (0, 0, 0) to ensure you don't loose accuracy. Also I suggest you examine the surface meshes prior to forcing the volume meshing. For more details review the comment "Mesh Diagnostics."
  2. The File->Preference, Physics->OpenFOAM->Parallel preference should only be set to Cluster if you are running on a Windows HPC Server 2008-based cluster with the Microsoft HPC Cluster Manager. Otherwise I recommend the Shared Memory option. With an 8 core processor I suggest you set the Parallel::Shared Memory->Processes preference to 7, leaving a core for Caedium. I'd also suggest experimenting with trying 25, but I don't think you are likely to see any better performance.

    If you change the mesh, often you change the simulation solving behavior - as you are seeing - finer meshes typically take longer to run. Also you might want to stop the simulation once the forces are converged rather than running to a fixed number of iterations.

    The Caedium licenses are not dependent on the platform (desktop, cluster, etc.), so whatever satisfies your functional requirements is your only consideration.

  3. There is no convergence critieria in Caedium. The simulations run for as many iterations as you specify, or until you pause the simulation.
  4. The Stop button is like an emergency stop. If you want to pause the simulation then toggle the Run button instead. If you do interrupt a simulation then you need to restore normal operations by toggling the Stop button. For more details review "Run, Stop, and Lock Controls".
  5. You'll have to create a new variable to normalize the Vector Variables->F (Force). For more details review the comment "Lift and Drag Coefficients."

Thank you

Submitted by Magrn on January 31, 2011 - 11:56.

1. The two meshes were the exact same thing, but havnt had that problem since.

2. As you suspected, using 7 and 25 processors seems same.

3. I am an experienced 'Fluent' user. So, was looking for a convergence criteria just like in most other solvers. Thought I missed it.

4. Getting used to these features.

5. ok.

Thanks for the elaborate answers and the links.

Feedback

Questions? Ideas? Problems?

Forums

Newsletter

more