Optimizing transient simulations

[mohd-afeef-badri]

For our transient simulations we unnecessarily reassemble the FEM matrix $A$ at each time step, this is unwanted for many simulations (at least for the ones that are in ArcaneFEM for now). We should

• assemble the matrix sparsity just onces as our simulations are not moving mesh ones
• let the choice if one needs to assemble the elements of matrix each time
• same choice could be given for RHS vector $b$

Currently in ArcaneFEM, the modules affected by this will be

• soildynamics
• elastodynamics
• heat

[mohd-afeef-badri]

we see here timings from simple soildynamics simulation with 15K DOF solved 30 times (time-steps), arcane is 10x slower than PSD.

Even parallel scaling is not optimal, although this might be due to simple preconditioner.

[mohd-afeef-badri]

data-time.txt

plot.txt

To get the image that I attached here, dowload the two txt files and run gnuplot plot.txt -p

[grospelliergilles]

Yes we definitively need to have a way to not recreate the matrix at each step. I do not know if it is easy to do. I will look at it as soon as possible

[mohd-afeef-badri]

With all the improvements in transient simulation

• not redoing the coefficients at every time-step
• making the sparsity just once now the code performs well