This is an implementation of k-means clustering algorithm using GPU acceleration with CUDA C++.
It is made for execution time comparison wrt CPU-only sequential version. For this reason the algorithm will not end when reaching convergence, but you need to specify the number of iterations when launching the program.
You can generate an N points dataset using datasetgen.py. You have to write also the number of clusters K and the standard deviation. The command will be like:
python datasetgen.py N K STD
example:
python datasetgen.py 1000 3 0.45
The code is made to work on 3 axis but the script will generate points with the third coordinate equal to 0.0. This is to ease the result checking with plot.py
./Kmeans N K I
Where N is the number of points to read from dataset, K is the number of clusters and I is the number of iterations
You can check output results with
python plot.py
Theese are results obtained using an NVIDIA GeForce GTX 980 Ti
| number of points | sequential (s) | CUDA (s) | speed up |
|---|---|---|---|
| 10 | 0.001 | 0.116 | x0.008 |
| 102 | 0.015 | 0.117 | x0.1 |
| 103 | 0.180 | 0.119 | x1.5 |
| 104 | 1.673 | 0.147 | x11.4 |
| 105 | 8.424 | 0.579 | x14.5 |
| 106 | 83.024 | 5.706 | x14.5 |
| 107 | 804.611 | 54.319 | x14.8 |
- CPU-only sequential version
- there is also a distributed system version
Parallel Computing - Computer Engineering Master Degree @University of Florence.