Forum Message

We have moved the forum to https://forum.ngsolve.org . This is an archived version of the topics until 05/05/23. All the topics were moved to the new forum and conversations can be continued there. This forum is just kept as legacy to not invalidate old links. If you want to continue a conversation just look for the topic in the new forum.

Notice

The forum is in read only mode.

any suggestions to speed up explicit DG updates?

Guosheng Fu
Topic Author
Offline
Elite Member

3 years 9 months ago #2909 by Guosheng Fu

any suggestions to speed up explicit DG updates? was created by Guosheng Fu

Dear all,

I have attached a demo code for incompressible two phase flow using hybrid-mixed FEM (RT0) for pressure and DG-P0 for saturation.
Currently, I use the standard DG updates for the saturation conservation law equation.
I believe this default version could be significantly improved, but I don't know what's the best option available now in ngsolve. I would like to get ideas about speeding up this code. Many thanks as always!

Best,
Guosheng

Attachments:

..._compile.ipynb

mneunteufel
Offline
Premium Member

3 years 9 months ago #2923 by mneunteufel

Replied by mneunteufel on topic any suggestions to speed up explicit DG updates?

Hi Guosheng,

I tried to modify your ForwardEuler by replacing the Matrix-Vector multiplication, but this did not really help.

Code:

def ForwardEuler(sh, dt): 
        rhsS.data = aS.mat * sh
        #rhsS.data = invm*rhsS
        fesS.SolveM(rhsS)
        return sh-dt*rhsS

Also to "inline" it did not bring a lot

Code:

#gfuS.vec.data = ForwardEuler(gfuS.vec, dt)
aS.Apply(gfuS.vec,rhsS)
gfuS.vec.data -= dt*invm*rhsS

However, the Compile flag made a huge difference! It is important to compile everything, not only a part of it

Code:

if kdg>0:
    aS += (-uh*fw(ss)*grad(tt)).Compile(true_compile, wait=True)*dx
...
aS += (uh*n*Fhat*tt).Compile(true_compile, wait=True)*dx(element_boundary=True)

With this I got a speed up of a factor 3!

It is also possible to speed up your "fast" part even more. Instead of calling mat.Invert every time you can "update" the inverse:

Code:

a.Assemble()
inv = a.mat.Inverse(fesP.FreeDofs(True),inverse="sparsecholesky")

while  time < tend-dt/2 :
            if stepS%nsubcycles == 0:
                a.Assemble()
                inv.Update()
                ...
                gfuP.vec.data = inv * rhsP

Then the symbolic factorization is done only once and not every step. This improved the "fast" part from 2.4 to 1.5 seconds on my machine.

Best
Michael

Guosheng Fu
Topic Author
Offline
Elite Member

3 years 9 months ago #2925 by Guosheng Fu

Replied by Guosheng Fu on topic any suggestions to speed up explicit DG updates?

Hi Michael,

Thanks about the two tricks, they are helpful.
I also get roughly a factor of 3 speedup with the correct use of the Compile flag, and a factor of 2 for the sparse Cholesky inverse update.

Changing element_boundary formulation to skeleton-based formulation further helps to speed-up the explicit DG solver slightly.

I am now thinking about using implicit DG solver for saturation. But Newton solver fails since .Other() only works for proxyfunctions, not gridfunctions. Do you have any suggestions on this issue?

Best,
Guosheng

Time to create page: 0.156 seconds

Login

NGSolve on Twitter