[AeroSonic]

Hi guys,

I have a problem with MATLAB, I am writing a compressible NV solver Using FEM (GLS), everything is fine, but I have a problem with the running time.

the running time for a one-time step ( using 90x30 grid ) is 22 seconds, 19 seconds of them are wasted in one line which is the inverse of the global matrix.

you can find Matlab code analyzer in attachments, I really don't know why this line takes this much time ?!!

I tried different MATLAB Versions ( 15b, 17b ) but it taking more time than the current ( 18b ).

I am going to change the code to FORTRAN soon but currently, I want to know the reason for this problem.

My specs:

1- Ryzen 5 2600.
2- Ram 3000 MHZ 2x8 Gb.
3- SSD

[praveen]

Matlab is slow if you use loops. If you cast everything as matrix-vector operations it can do better.

For CFD, Matlab is not so great, ok for some initial learning, but you soon reach its limitations.

Do you have to assemble your matrices every time step ?

You are using a direct solver and it has to do a LU decomposition every time.

If matrix does not change you can perhaps optimise, perform LU decomposition once,

https://www.mathworks.com/help/matlab/ref/lu.html

and reuse it.

[AeroSonic]

Originally Posted by praveen

Matlab is slow if you use loops. If you cast everything as matrix-vector operations it can do better.

For CFD, Matlab is not so great, ok for some initial learning, but you soon reach its limitations.

Do you have to assemble your matrices every time step ?

You are using a direct solver and it has to do a LU decomposition every time.

If matrix does not change you can perhaps optimise, perform LU decomposition once,

https://www.mathworks.com/help/matlab/ref/lu.html

and reuse it.

I am going to learn Fortran soon but.

as this is a compressible viscous and unsteady, yes i am constructing the matrices each time step and they are changing from element to element.

every time step, the code does the inverse line. I am using the mldivide.

[praveen]

In your case, most time is taken in solve phase. Your matrix is not that big so the direct solvers should be fine. Maybe this is normal. The solve phase will be the most expensive in such a method.

[FMDenaro]

Could you show a sketch at a couple of time steps of the sparsity of the matrix?
However, in any CFD code a solver for algebric system takes the most part of computational time.

[AeroSonic]

Originally Posted by FMDenaro

Could you show a sketch at a couple of time steps of the sparsity of the matrix?
However, in any CFD code a solver for algebric system takes the most part of computational time.

It was a problem in matlab, I have an AMD processor and matlab isnt configured for it, someone helped me

https://www.mathworks.com/matlabcent...#answer_401963