search

barbagroup / inexact-gmres

Paper: inexact GMRES with fast multipole method and low-p relaxation
9 stars 4 forks source link

why is speedup of Stokes not higher, even? #2

Closed labarba closed 2 years ago

labarba commented 9 years ago

The residual history of the test with Stokes flow shows that it's reaching the solver tolerance after ~27 iterations. This is 8k panels, and speedup of the relaxation strategy is 3.5x.

For the Laplace equation, forcing the iterations up to 25 resulted in a speed-up of 3.2x (table 3.4 in the paper, 5.4 in the thesis), when using 32k panels. It looks like there wasn't much difference in the speedup whether you used 8k or 32k panels.

Since the Stokes kernel expansions are equivalent to four Laplace kernels, I am confused as to why the speedups are not even higher for Stokes. It seems to be performing about the same for both kernels, following the reasoning above.

slayton58 commented 9 years ago

Remember that the local evaluation for stokes is ~9x (from memory) that of Laplace as well (generating a 3x3 matrix instead of a scalar value).