cuIBM
A GPU-based Immersed Boundary Method code
cuIBM - A GPU-based Immersed Boundary Method code

cuIBM solves the 2D incompressible Navier-Stokes equations with an immersed-boundary method on a single CUDA-capable GPU device with the CUSP library.

Currently, cuIBM runs only on Unix-based systems (no support on Windows) and was last tested on Ubuntu 16.04.


Documentation

User's documentation is available on the Wiki pages of the cuIBM repository.

Doxygen API documentation is available here.


Installation

Dependencies (last tested)

GNU C++ Compiler

Install g++ using the following command:

> sudo apt-get install g++

Check the version of g++ installed:

> g++ --version

Other development and version control tools can be installed by following the instructions under Step 1 in the CompilingEasyHowTo page on the Ubuntu Community Help Wiki. Software developers will find it useful to install all of them.

NVIDIA's CUDA Compiler

Download and install the CUDA Toolkit.

Check the version of NVCC installed:

> nvcc --version

CUSP Library

CUSP is a library that provides routines to perform sparse linear algebra computations on Graphics Processing Units. It is written using the CUDA programming language and runs code on compatible NVIDIA GPUs.

CUSP is currently hosted on GitHub. cuIBM has been tested and works with version 0.5.1, available for download here.

> export CUSP_DIR=$HOME/software/cusp/0.5.1
> mkdir -p $CUSP_DIR
> wget https://github.com/cusplibrary/cusplibrary/archive/v0.5.1.tar.gz
> tar -xvf v0.5.1.tar.gz -C $CUSP_DIR --strip-components=1
> rm -f v0.5.1.tar.gz

Boost library

In cuIBM, we use the parser YAML-CPP (version 0.5.1 bundled in cuIBM) that requires header files from the Boost library.

> export BOOST_DIR=$HOME/software/boost/1.64.0
> mkdir -p $BOOST_DIR
> wget https://dl.bintray.com/boostorg/release/1.64.0/source/boost_1_64_0.tar.gz
> tar -xvf boost_1_64_0.tar.gz -C $BOOST_DIR --strip-components=1
> rm -f boost_1_64_0.tar.gz

Compiling cuIBM

Clone cuIBM:

> mkdir -p $HOME/software
> cd $HOME/software
> git clone https://github.com/barbagroup/cuIBM.git

To compile cuIBM, make sure you have set correctly the environment variables CUSP_DIR and BOOST_DIR to point to their respective folder:

> export CUSP_DIR=$HOME/software/cusp/0.5.1
> export BOOST_DIR=$HOME/software/boost/1.64.0

We also recommend setting the environment variable CUIBM_DIR to point to the location of the cuIBM folder. While the code can be compiled and run without setting this variable, some of the validation scripts provided make use of it.

> export CUIBM_DIR=$HOME/software/cuIBM

To compile cuIBM, navigate to the cuIBM directory:

> cd $HOME/software/cuIBM

or

> cd $CUIBM_DIR

then, run make:

> make

Note: the current Makefile compiles cuIBM with the flag -arch=sm_35 (compute capability of 3.5). You should adapt the Makefile if your NVidia card has a different compute capability.

Run an example (2D flow around cylinder at Re=40) to check that cuIBM has been built correctly:

> bin/cuibm -directory examples/cylinder/Re40

Finally, you can add cuIBM to your path:

> export PATH="$HOME/software/cuIBM/bin:$PATH"

Users-documentation is available in the Wiki pages of the GitHub repository.

Post-processing

To post-process the numerical solution from cuIBM, we provide Python scripts (present in each case in the folder examples). They make use of the package snake that is bundled in the external folder of cuIBM (version 0.3).

To install snake:

> cd $CUIBM_DIR/external/snake-0.3
> python setup.py install

snake requires Python (2.7 or 3.5), Matplotlib, Scipy, and Pandas.


Example: Flow over impulsively started cylinder (Re=550)

To run the example:

> cd $CUIBM_DIR/examples/cylinder/Re550
> cuibm

To plot the instantaneous drag coefficient and the vorticity field at saved time steps:

> python scripts/plotDragCoefficient.py
> python scripts/plotVorticity.py

Figures are saved in the folder images of the simulation directory.

We also provide a detailed description for several examples:

with comparison to experimental and computational results from other studies.


List of examples

To run any of the examples listed above:

> cd $CUIBM_DIR
> make <examplename>

or change directory to the corresponding example's folder and

> cuibm

The biggest cases (for the gliding snake) requires a GPU device with at least 4GB of memory.

To test an example on a smaller number of time steps, you can change the value of the parameters nt in the input file simParams.yaml of the example directory. (Note that the flow solution is saved every nsave time steps which value can be changed in the same input file.)


Papers published using cuIBM

Contact

Please e-mail Anush Krishnan or Olivier Mesnard, if you have any questions, suggestions or feedback.

To report bugs, please use the GitHub issue tracking system. We are also open to pull-requests.