Out of memory during LU factorization

Please login with a confirmed email address before reporting spam

Could you help to understand what is wrong in my model? I get not converged solution error. You can find the model in attachment



9 Replies Last Post Oct 10, 2024, 12:21 p.m. EDT
Robert Koslover Certified Consultant

Please login with a confirmed email address before reporting spam

Posted: 3 weeks ago Sep 27, 2024, 12:55 p.m. EDT

Looks like you are simply running out of memory. There are many ways to address this. See https://www.comsol.com/support/knowledgebase/830 and https://www.comsol.com/support/knowledgebase/1030

-------------------
Scientific Applications & Research Associates (SARA) Inc.
www.comsol.com/partners-consultants/certified-consultants/sara
Looks like you are simply running out of memory. There are many ways to address this. See https://www.comsol.com/support/knowledgebase/830 and https://www.comsol.com/support/knowledgebase/1030

Please login with a confirmed email address before reporting spam

Posted: 3 weeks ago Sep 30, 2024, 10:29 a.m. EDT
Updated: 3 weeks ago Sep 30, 2024, 11:08 a.m. EDT

Hi, my mesh is composed by about 1400000 tetrahedral elements (i suppose they are linear tetrahedral, let me know if there is a way to check). I have 4 dependent variables (electric field components and 1 S-parameter), so the number of degrees of freedom should be 0.2*1400000*4= 1200000 more or less. So less than 5 Gb of Physical memory should be enough according to comsol documentation (https://www.comsol.it/blogs/much-memory-needed-solve-large-comsol-models). The Log of my simulation says this

Compile Equations: Frequency Domain in Study 1/Solution 1 (sol1) ---------
Started at Sep 30, 2024, 4:10:09 PM.
Geometry shape function: Quadratic Lagrange
Running on 4 x Intel(R) Xeon(R) Gold 6154 CPU at 3.00 GHz.
Using 4 sockets with 2 cores in total on rat4.rfx.local.
Available memory: 773.70 GB.
Canceled
Time: 9 s.
Physical memory: 6.36 GB
Virtual memory: 21.79 GB
Ended at Sep 30, 2024, 4:10:18 PM.
----- Compile Equations: Frequency Domain in Study 1/Solution 1 (sol1)

So it looks like it is using more than 6 GB and still is going out of memory. Why?

Hi, my mesh is composed by about 1400000 tetrahedral elements (i suppose they are linear tetrahedral, let me know if there is a way to check). I have 4 dependent variables (electric field components and 1 S-parameter), so the number of degrees of freedom should be 0.2\*1400000\*4= 1200000 more or less. So less than 5 Gb of Physical memory should be enough according to comsol documentation (https://www.comsol.it/blogs/much-memory-needed-solve-large-comsol-models). The Log of my simulation says this Compile Equations: Frequency Domain in Study 1/Solution 1 (sol1) --------- Started at Sep 30, 2024, 4:10:09 PM. Geometry shape function: Quadratic Lagrange Running on 4 x Intel(R) Xeon(R) Gold 6154 CPU at 3.00 GHz. Using 4 sockets with 2 cores in total on rat4.rfx.local. Available memory: 773.70 GB. Canceled Time: 9 s. Physical memory: 6.36 GB Virtual memory: 21.79 GB Ended at Sep 30, 2024, 4:10:18 PM. ----- Compile Equations: Frequency Domain in Study 1/Solution 1 (sol1) So it looks like it is using more than 6 GB and still is going out of memory. Why?

Henrik Sönnerlind COMSOL Employee

Please login with a confirmed email address before reporting spam

Posted: 3 weeks ago Sep 30, 2024, 11:11 a.m. EDT

As you can see in the attached printout, you are using second-order elements. So the DOF count is much larger. The number of DOF is printed in the Messages tab when you start the analysis.

You select the order of the elements in the Discretization section in the settings for the physics interface.

-------------------
Henrik Sönnerlind
COMSOL
As you can see in the attached printout, you are using second-order elements. So the DOF count is much larger. The number of DOF is printed in the Messages tab when you start the analysis. You select the order of the elements in the Discretization section in the settings for the physics interface.

Edgar J. Kaiser Certified Consultant

Please login with a confirmed email address before reporting spam

Posted: 3 weeks ago Sep 30, 2024, 11:31 a.m. EDT

With linear discretization it takes about 32 GB and it completes. The mesh may be finer than needed. You can save a lot of memory by adjusting the mesh.

-------------------
Edgar J. Kaiser
emPhys Physical Technology
www.emphys.com
With linear discretization it takes about 32 GB and it completes. The mesh may be finer than needed. You can save a lot of memory by adjusting the mesh.

Please login with a confirmed email address before reporting spam

Posted: 2 weeks ago Oct 1, 2024, 10:33 a.m. EDT

Ok thank you for the help. I have another doubt, as you can see my model is perfectly symmetryc with respect to the zx plane, so I wonder If can simulate only half of my geometry to save computational resources and time. I already did this in other models, but the problem now is that on the symmetry plane my fields (the fields of a input TE10 mode) are both tangent to the plane, so I can't use PMC or PEC boundary conditions, rigth?. What boundary condition would you suggest me to use in this case?

Thank you in advance

Ok thank you for the help. I have another doubt, as you can see my model is perfectly symmetryc with respect to the zx plane, so I wonder If can simulate only half of my geometry to save computational resources and time. I already did this in other models, but the problem now is that on the symmetry plane my fields (the fields of a input TE10 mode) are both tangent to the plane, so I can't use PMC or PEC boundary conditions, rigth?. What boundary condition would you suggest me to use in this case? Thank you in advance

Edgar J. Kaiser Certified Consultant

Please login with a confirmed email address before reporting spam

Posted: 2 weeks ago Oct 1, 2024, 11:48 a.m. EDT

The E-field is perpendicular to the symmetry plane, no?

-------------------
Edgar J. Kaiser
emPhys Physical Technology
www.emphys.com
The E-field is perpendicular to the symmetry plane, no?

Please login with a confirmed email address before reporting spam

Posted: 1 week ago Oct 10, 2024, 6:37 a.m. EDT

Yes you are right, my mistake. Now I have to do a bigger simulation where I can't cut my domain in two because I have to excite a combination of different modes at the antenna port, and so I have a very large DOF number (about 40000000). Which solver would you suggest to use when having such high DOF number? I already switched to linear discretization, and I can't further reduce my mesh accuracy otherwise I won't catch the physics of the problem (radiation pattern of antenna with mode converted input at frequency >80 GHz). A strategy could be to try to use hexahedral mesh, but it looks very complicated with my geometry. I have about 90 GB at maximum to use on the server where I am working. Could an iterative solver be better than PARDISO?

Yes you are right, my mistake. Now I have to do a bigger simulation where I can't cut my domain in two because I have to excite a combination of different modes at the antenna port, and so I have a very large DOF number (about 40000000). Which solver would you suggest to use when having such high DOF number? I already switched to linear discretization, and I can't further reduce my mesh accuracy otherwise I won't catch the physics of the problem (radiation pattern of antenna with mode converted input at frequency >80 GHz). A strategy could be to try to use hexahedral mesh, but it looks very complicated with my geometry. I have about 90 GB at maximum to use on the server where I am working. Could an iterative solver be better than PARDISO?

Henrik Sönnerlind COMSOL Employee

Please login with a confirmed email address before reporting spam

Posted: 1 week ago Oct 10, 2024, 8:19 a.m. EDT

Using an iterative solver is the only possibility you have for solving very large problems.

Have you considered the possibility to split the problem into one symmetric and one antisymmetric using different sets of boundary conditions? Any linear problem on a symmetric geometry can be considered as a superposition of a symmetric and an antisymmetric solution.

-------------------
Henrik Sönnerlind
COMSOL
Using an iterative solver is the only possibility you have for solving very large problems. Have you considered the possibility to split the problem into one symmetric and one antisymmetric using different sets of boundary conditions? Any linear problem on a symmetric geometry can be considered as a superposition of a symmetric and an antisymmetric solution.

Edgar J. Kaiser Certified Consultant

Please login with a confirmed email address before reporting spam

Posted: 1 week ago Oct 10, 2024, 12:21 p.m. EDT

Henrik's suggestion of superposing a symmetric and antisymmetric solution worked for me in very big acoustic models. These models were geometrically symmetric but the wave excitation was not. It saves half the memory and also half the solution time.

-------------------
Edgar J. Kaiser
emPhys Physical Technology
www.emphys.com
Henrik's suggestion of superposing a symmetric and antisymmetric solution worked for me in very big acoustic models. These models were geometrically symmetric but the wave excitation was not. It saves half the memory and also half the solution time.

Reply

Please read the discussion forum rules before posting.

Please log in to post a reply.

Note that while COMSOL employees may participate in the discussion forum, COMSOL® software users who are on-subscription should submit their questions via the Support Center for a more comprehensive response from the Technical Support team.