███    ███ ██   ██ ██ ████████ ██████   ██████  
████  ████ ██   ██ ██    ██         ██ ██       
██ ████ ██ ███████ ██    ██     █████  ███████  
██  ██  ██ ██   ██ ██    ██         ██ ██    ██ 
██      ██ ██   ██ ██    ██    ██████   ██████         

GPU-based Finite difference code for DNS of Multiphase Homogenous isotropic turbulence (PFM and particles)

Main developer: A. Roccon

Tested on:

• Milton (1 x RTX5000)
• Leonardo (1 x A100)

Current capabiltiies:

• DNS of single-phase flow
• Phase-field method (ACDI method)
• Tracking of tracers
• Paraview files for particles
• Surface tension forces
• Euler/AB2 for time integration (default AB2)

Future developments:

• Tracking of inertial particles
• FFTW for CPU debug run
• Memory optimization

Performance (NS only)

• 128 x 128 x 128 | RTX5000@milton | 14 ms/timestep
• 128 x 128 x 128 | A100@Leonardo | 5 ms/timestep
• 256 x 256 x 256 | A100@Leonardo | 40 ms/timestep
• 512 x 512 x 512 | A100@Leonardo | 320 ms/timestep

Performance (NS + PFM)

• 128 x 128 x 128 | RTX5000@milton | 30 ms/timestep

Max resolution tested:

• 256 x 256 x 256 | RTX5000@milton - 16 GB VRAM
• 768 x 768 x 768 | A100@Leonardo - 64 GB VRAm

Systems supported:

• Unix + nvfortran

To run a simulation:

• go to src folder and run ./compile_local.sh (if you have one GPU, UNIX system) or ./compile_Leo.sh to run on Leondaro supercomputer.

Parallelization strategy

• The code is serial and exploit a single GPU (GPU-resident)

Output files.

• Files containing the Eulerian fields (u_***, v_****, w_****, p_**** and phi_****) and the particle positions (xp_***) are stored in src/output