CPU Time for many load conditions
CPU Time for many load conditions
(OP)
Hi,
I have a big model (830000 nodes) with 12 subcases for linear static. When I ran this model with only one subcase it took about 3.5 hours on our server. Now, I am looking for a simple rule of thumb for the following ratio?
(CPU time needed for N subcases) / (CPU time needed for only one subcase)
I think that the above ratio must be less than N. Please let me know about your experience.
Regards,
AAY
I have a big model (830000 nodes) with 12 subcases for linear static. When I ran this model with only one subcase it took about 3.5 hours on our server. Now, I am looking for a simple rule of thumb for the following ratio?
(CPU time needed for N subcases) / (CPU time needed for only one subcase)
I think that the above ratio must be less than N. Please let me know about your experience.
Regards,
AAY





RE: CPU Time for many load conditions
Should be that most of the time is spent solving for the stiffness matrix. If you have only one boundary condition and 12 different forces for you 12 subcases, this should barely take longer than N since there is only 1 stiffness matrix to solve for. But, if you have 12 different boundary conditions, there should be a different stiffness matrix for each case and the time should be about 12N.
RE: CPU Time for many load conditions
RE: CPU Time for many load conditions
solve time = DOF**(2-4)
This is only an approximate relationship, which requires tuning of the exponent.
Cheers,
-- drej --
RE: CPU Time for many load conditions
I think that the most important part of a FEA solving time is related to the global stiffness matrix. This matrix is affected only by the geometrical and mechanical properties of model (boundary conditions don't affect this matrix). Well, I think that by running a model with N sub-cases, I should save significantly CPU time. Because, it doesn't need to build the global stiffness matrix again and again. It builds this matrix only one time. What do you think?
Regards,
AAY
http://www.geocities.com/fea_tek/asd/
RE: CPU Time for many load conditions
A typical linear FEA run consists of a number of phases, as follows:
a) Read the model data, check, set up stiffness matrix, etc. This phase should be pretty much a constant for any given model, regardless of the number of load cases.
b) Read the load data, check, set up load vectors. This phase will be roughly proportional to the number of load cases (assuming all load cases are roughly similar in number of node and element loads, overall load case complexity, etc).
c) Reduction and back-substitution. As drej notes, this phase is roughly proportional to DOF^(2 to 4), but should be pretty much a constant for any given model, regardless of the number of load cases.
d) Calculating nodal deflections, element stresses, nodal reactions, etc. This phase will be roughly proportional to the number of load cases.
Of course, other factors enter the equation – in particular, do you have sufficient RAM to solve the equations substantially in RAM, or is the computer constantly swapping large amounts of data to and from the hard disc.
If you can review the log file for your first run, you may be able to work out roughly how much time your run spent on each of phases a) to d). Then your estimate for the complete run time for a larger number of load cases would be:
Run Time = (a + c) + (No. of Load Cases) * (b + d)
For “small” problems, a, b and d can actually take longer than the reduction and back-substation phase (c). For “large” models, the reduction and back-substitution phase will dominate. (I would consider a model with 830,000 nodes to be a “large” model in this context.)
(Your mileage may vary!)
RE: CPU Time for many load conditions
AAY