https://github.com/stargate01/hpcgol
High performance Conway's Game of Life
https://github.com/stargate01/hpcgol
conways-game-of-life hpc mpi openmp simd vectorization
Last synced: 6 months ago
JSON representation
High performance Conway's Game of Life
- Host: GitHub
- URL: https://github.com/stargate01/hpcgol
- Owner: StarGate01
- License: mit
- Created: 2019-05-03T16:48:40.000Z (over 6 years ago)
- Default Branch: master
- Last Pushed: 2019-05-24T12:34:32.000Z (over 6 years ago)
- Last Synced: 2024-05-01T14:11:43.811Z (over 1 year ago)
- Topics: conways-game-of-life, hpc, mpi, openmp, simd, vectorization
- Language: Fortran
- Size: 4.99 MB
- Stars: 9
- Watchers: 3
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# High performance Conway's Game of Life
High performance Conway's Game of Life. See https://en.wikipedia.org/wiki/Conway%27s_Game_of_Life.
## Modules
- Exploitation of cache locality and memory layout (`simple`)
- Optimized kernel (`simple_opt`)
- Vectorization (sub-optimal in this use-case, not recommended) using
- The Auto-Vectorizer (`simple_simd`)
- OpenMP SIMD (`simple_simd2`)
- Thread-level parallelism using OpenMP (`threads`)
- CPU-Node parallelism using MPI (`nodes`)
- Thread-level and CPU-Node hybrid parallelism using OpenMP and MPI (`hybrid`)
## Installation
Run `sudo make deps` if your package manager ist apt, otherwise install these packages:
- gfortran
- mpich
- libmpich-dev
## Compilation
Run `make all` or `make modulename` or `make bin/modulename` to build all resp. a single program.
Run `make clean` to remove all binaries.
### Debugging and optimization
To enable debug symbols and disable code optimization, define the `DEBUG` makefile variable when calling `make`, like this: `make modulename DEBUG=1`.
To enable all possible compiler optimizations in order to produce fast code, do not define the `DEBUG` variable. Note that due to aggressive optimization binaries are generally not portable between different CPUs.
### MPICH2 vs OpenMPI
If you want to use `OpenMPI` instead of `MPICH2`, the code should compile and work but I have not tested this yet but may support it in the future.
### GNU compiler vs Intel compiler
Instead of the GNU compiler `gfortran` the intel compiler `ifort` may be used by defining the makefile variable `INTEL`. In general, every OpenMP- and MPI-capable Fortran-2008 standard compliant compiler should work.
## Execution
The included makefile has a `run` target, which should be used to launch the binaries. Note that also invokes the compilation. The syntax and some default values are as follows:
`make run EXE=modulename ARG_PRINT=1 ARG_STEPS=15 ARG_WIDTH=10 ARG_HEIGHT=10 ARG_THREADS=1 ARG_NODES=1`
- `EXE`: The name of the module to launch, see above.
- `ARG_PRINT`: 1 to print the field as ACII, 0 to not print it. This is only recommended for small field sizes.
- `ARG_STEPS`: The number of iterations to compute.
- `ARG_WIDTH`: The width in cells of the field.
- `ARG_HEIGHT`: The height in cells of the field.
- `ARG_THREADS`: The number of threads to use. Note that more threads than your system has CPU cores or hyper-threads may slow things down.
- `ARG_NODES`: The number of MPI nodes to use. If there is no MPI cluster available, `hydra` will automatically emulate a cluster using multiple processes.
### Recommended settings
Use the default values above to test the program using its graphical output.
A field size of `10000` times `10000` is a reasonable first scale test, it runs with about 0.2 - 1 steps per second on most laptops and PCs, depending on optimization and CPU speed.
### Memory requirements
The minimum amount of required RAM is roughly `2 * ARG_WIDTH * ARG_HEIGHT` bytes. When running on a MPI cluster, this memory is equally distributed.
## IDE support
The Visual Studio Code IDE (See https://code.visualstudio.com/) is supported. But any other IDE capable of launching makefiles should work.