Title: "Experience report of physics-informed neural networks in fluid simulations: pitfalls and frustration" Preprint on arXiv: https://arxiv.org/abs/2205.14249
Cases under folder petibm were run with the Singularity image created by petibm-0.5.4rc2-hpcx207-cuda102.singularity under singularity folder.
Cases under folder modulus were run with the Singularity image created by modulus-22.03.singularity under singularity folder.
However, Modulus is not an open-source software.
You need to go to NVIDIA's developer zone, download Modulus 22.03, and follow the instruction to build a Docker image in the local registry.
Then, you can use modulus-22.03.singularity to create the Singularity image.
(Update: at the time when we ran these cases, we needed to create the Docker image manually.
However, Modulus has provided a Docker image on the NGC platform, so there's no need to manually build it now.)
PetIBM
Basically just go into each case folder and run
$ CUDA_VISIBLE_DEVICES=<list of target gpus> \
mpiexec \
-n <number of CPUs> \
singularity exec --nv <petibm singularity image> petibm-navierstokesThe results shown in the paper were obtained using 1 single K40c GPU and 6 CPU cores of i7-5930K.
For cylinder flow, replace petibm-navierstokes with petibm-decoupledibpm.
Modulus
Each case has job.sh for Slurm scheduler.
Modify the script for your target cluster, and then do sbatch job.sh to submit each job one by one.
The original resource used (the partition shown in job.sh) was a node of NVIDIA DGX-A100-640G, but it should work on other GPUs.
Just note the memory usage of some cases may be non-trivial, and probably only A100 (both 40GB and 80GB variants) can host all runs.
V100 32GB may be able to handle most cases except some extreme ones.
Post-processing
First, generate processed data by executing the following three python scripts: modulus_cylinder_2d_re200.py, modulus_tgv_2d_re100.py, petibm_tgv_2d_re100.py.
Then generate figures using other scripts.
Generated figures will be in a folder called figures.