Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/dizys/nyu-gpu-final-project
NYU CV Final Project: Comparing CUDA to OpenMP on GPUs
https://github.com/dizys/nyu-gpu-final-project
Last synced: 15 days ago
JSON representation
NYU CV Final Project: Comparing CUDA to OpenMP on GPUs
- Host: GitHub
- URL: https://github.com/dizys/nyu-gpu-final-project
- Owner: dizys
- License: mit
- Created: 2022-11-17T20:13:35.000Z (about 2 years ago)
- Default Branch: main
- Last Pushed: 2022-12-07T21:28:06.000Z (about 2 years ago)
- Last Synced: 2024-12-20T20:03:10.607Z (21 days ago)
- Language: C++
- Homepage:
- Size: 137 KB
- Stars: 1
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# nyu-gpu-final-project
Comparing CUDA to OpenMP on GPUs with K-Means, BFS, Floyd-Warshall, and N-Queens algorithms.
## Getting Started
### Setup CIMS GPU machines
We experimented on `cuda3.cims.nyu.edu`.
1. Load the CUDA module:
```bash
module load cuda-10.1
```
2. Download GCC with OpenMP GPU-offloading support:
```bash
mkdir -p /tmp/$(whoami)
wget -O /tmp/$(whoami)/gcc.zip https://github.com/nyu-multicore/cims-gpu/releases/download/gcc/gcc-11.3.1_cims_gpu_offload_22112501.zip && unzip /tmp/$(whoami)/gcc.zip -d /tmp/$(whoami) && rm -f /tmp/$(whoami)/gcc.zip
```
> **Or for step 2, alternatively, you can build GCC from source with OpenMP GPU-offload support**
>
> ```bash
> wget -O build_gcc.sh https://gist.githubusercontent.com/dizys/8dedbe94439b91d759b6c1e6e316d542/raw/3ddbd8def8cc5bc7ce42549317820df16daf9e96/build_gcc_with_offload.sh && sh build_gcc.sh && rm -f build_gcc.sh
> ```
>
> This will take roughly 30 minutes to build GCC 11 from source. And GCC will be installed to `/tmp//gcc` temporarily.
### Compile the programs
Before compilation, environment variable `LD_LIBRARY_PATH` must be set to the path of the GCC installation.
```bash
export MY_GCC_PATH=/tmp/$(whoami)/gcc
export LD_LIBRARY_PATH=$MY_GCC_PATH/lib64:${LD_LIBRARY_PATH}
```
Then, build the project:
```bash
make GPP_BIN=$MY_GCC_PATH/bin/g++
```
### Download datasets
```bash
mkdir -p /tmp/$(whoami)/data
# KMeans datasets
wget -O /tmp/$(whoami)/data/kmeans_10000.txt https://media.githubusercontent.com/media/nyu-multicore/k-means/main/data/dataset-10000.txt
wget -O /tmp/$(whoami)/data/kmeans_100000.txt https://media.githubusercontent.com/media/nyu-multicore/k-means/main/data/dataset-100000.txt
wget -O /tmp/$(whoami)/data/kmeans_1000000.txt https://media.githubusercontent.com/media/nyu-multicore/k-means/main/data/dataset-1000000.txt
wget -O /tmp/$(whoami)/data/kmeans_5000000.txt https://media.githubusercontent.com/media/nyu-multicore/k-means/main/data/dataset-5000000.txt
wget -O /tmp/$(whoami)/data/kmeans_10000000.txt https://media.githubusercontent.com/media/nyu-multicore/k-means/main/data/dataset-10000000.txt
# BFS datasets
wget -O /tmp/$(whoami)/data/graph_g1000_s100.txt https://github.com/nyu-multicore/cims-gpu/releases/download/bfs-data/graph_g1000_s100.txt
wget -O /tmp/$(whoami)/data/graph_g2000_s100.txt https://github.com/nyu-multicore/cims-gpu/releases/download/bfs-data/graph_g2000_s100.txt
wget -O /tmp/$(whoami)/data/graph_g4000_s100.txt https://github.com/nyu-multicore/cims-gpu/releases/download/bfs-data/graph_g4000_s100.txt
wget -O /tmp/$(whoami)/data/graph_g8000_s100.txt https://github.com/nyu-multicore/cims-gpu/releases/download/bfs-data/graph_g8000_s100.txt
wget -O /tmp/$(whoami)/data/graph_g16000_s100.txt https://github.com/nyu-multicore/cims-gpu/releases/download/bfs-data/graph_g16000_s100.txt
# Floyd-Warshall datasets
cd bin && ./generategraph && cd .. # generate dataset .txt
# N-Queens: N-Queens programs don't need any extra dataset files to run
```
## Run the programs
```bash
cd bin
# KMeans
./kmeans_seq /tmp/$(whoami)/data/kmeans_.txt # run the sequential version
./kmeans_cuda /tmp/$(whoami)/data/kmeans_.txt # run the CUDA version
./kmeans_openmp /tmp/$(whoami)/data/kmeans_.txt # run the OpenMP version
# BFS
./bfs_seq /tmp/$(whoami)/data/graph_g_s.txt # run the sequential version
./bfs_cuda /tmp/$(whoami)/data/graph_g_s.txt # run the CUDA version
./bfs_openmp /tmp/$(whoami)/data/graph_g_s.txt # run the OpenMP version
# Floyd-Warshall
./warshall_seq .txt # run the sequential version
./warshall_cuda .txt # run the CUDA version
./warshall_openmp .txt # run the OpenMP version
# N-Queens
./nqueens_seq # run the sequential version
./nqueens_cuda # run the CUDA version
./nqueens_openmp # run the OpenMP version
```
## Experiment Results
We experimented on `cuda3.cims.nyu.edu`.
For time spent measurement, every experiment setting has been run 5 times and should be averaged to get the final result.
Result data files:
| Filename | Description |
| ------------------------------------ | ----------------------------------------------------------------------------------------------------- |
| [exp_data.csv](exp_data.csv) | Raw time cost experiment data with 5 iterations |
| [exp_data_avg.csv](exp_data_avg.csv) | Averaged time cost experiment data, generated by [exp_data_crunching.ipynb](exp_data_crunching.ipynb) |
| [exp_mem_data.csv](exp_mem_data.csv) | Max GPU memory usage data |
## License
This project is licensed under the MIT License - see the [LICENSE](LICENSE) file for details