The solving speeds of svSolver and svFSI in fluid problems

[Li Xia]

Hello,

I have recently been studying fluid problems. Both svSolver and svFSI can be used in fluid problems. svFSI is more standardized in code writing, more widely used and maintained, which has attracted my attention. Consequently, I wanted to compare the results of svSolver and svFSI in terms of speed and periodicity in fluid problems.

Under the same conditions, including the same grid, step size, number of cores, and boundary conditions, using the default solving method, svSolver and svFSI have shown consistent results in the first two cycles, with good periodicity. However, in terms of solving speed, svSolver took approximately 45 minutes for the first cycle and 25 minutes for the second, while svFSI took around 6 hours for both the first and second cycles, indicating a significant difference in speed.

I'm curious if the two sets of codes handle fluid problems in the same way. Can using Trilinos significantly accelerate the solving speed of svFSI?

Code: Select all

# Part of svFSI.inp
# Equations
Add equation: fluid {
   Coupled: 1
   Min iterations: 3
   Max iterations: 10
   Tolerance: 1e-3
   Backflow stabilization coefficient: 0.2

   Density: 1.06
   Viscosity: Constant {Value: 0.04}

   Output: Spatial {
      Velocity: t
      Pressure: t
      Traction: t
      WSS: t
   }

   LS type: NS
   {
      Max iterations:      10
      NS-GM max iterations: 3
      NS-CG max iterations: 500
      Tolerance:           1D-3
      NS-GM tolerance:     1D-3
      NS-CG tolerance:     1D-3
      Krylov space dimension: 50
   }

Best Regards,
Li Xia

[David Parker]

Hi Li Xia,

We have seen that svFSI and svSolver have about the same execution speed for CFD computations. You didn't give any information about mesh sizes, the number of cores used and the platform you used to make the comparison and how you did the timings, perhaps some dependency there.

The Trilinos preconditioners do improve performance for solid mechanics problems, less so for CFD I think.

Cheers,
Dave

[Li Xia]

Hi Dave,

I simulated the cerebral vascular system using svSolver and svFSI with the same grid consisting of 3.25 million cells on the Beijing Super Cloud Computing platform, utilizing 32 cores for testing. I estimated the simulation speed by comparing the times of the files generated by both (restart.xx.xx files for svSolver and vtu files for svFSI), and there is a significant difference in the time for each step, which can also be observed from the histor.dat file (refer to the two images below). The time step is 0.002, and the boundary condition is Resistance BC.

Best Regards,
Li Xia

[David Parker]

Hi Li Xia,

Your svFSI simulation is not converging in 10 iterations so your simulation is not setup correctly.

Cheers,
Dave