This project presents a comparative study of Graph Neural Network (GNN) acceleration techniques. We explore and evaluate two sparsification and two sampling methods:
- DropEdge
- NeuralSparse
- GraphSAGE
- GraphSAINT
To run any of the experiments, simply run the main.py file with the desired parameters. Specify the sparsification and sampling methods as command-line arguments with the --sparse and --sample flags, respectively. For example:
python main.py --sparse dropedge --sample graphsage
All available options can be found via the --help flag for main.py.
The methods are compared based on the following parameters:
- CPU Usage (%)
- GPU Usage (%)
- CPU Power Consumption (W)
- GPU Power Consumption (W)
- RAM Usage
- VRAM Usage
- Time per Epoch
- Number of Epochs
- ROC AUC
- EDP (Energy-Delay Product)
This study aims to provide insights into the efficiency and performance trade-offs of different GNN acceleration techniques.
The results of the comparative study are given in the following tables. A full discussion of these results can be found in the report accompanying this project.
TABLE I: System parameters while training
| Sparsification Method |
Sampling Method |
CPU Usage (%) |
CPU Power (W) |
GPU Usage (%) |
GPU Power (W) |
RAM (MB) |
vRAM (MB) |
|---|---|---|---|---|---|---|---|
| None | Uniformly Random | 29.6 | 10.2 | 6.3 | 4.3 | 4212 | 1450 |
| None | GraphSAGE | 40.9 | 18.6 | 15.6 | 5.0 | 6481 | 433 |
| None | GraphSAINT | 45.0 | 20.1 | 23.7 | 9.8 | 7121 | 1055 |
| DropEdge | Uniformly Random | 40.3 | 18.0 | 42.3 | 20.0 | 3141 | 3388 |
| DropEdge | GraphSAGE | 44.0 | 19.5 | 41.8 | 19.3 | 6778 | 2560 |
| DropEdge | GraphSAINT | 47.2 | 21.2 | 23.9 | 10.3 | 7391 | 849 |
| NeuralSparse | Uniformly Random | 36.5 | 14.2 | 27.4 | 16.4 | 3074 | 1639 |
| NeuralSparse | GraphSAGE | 38.9 | 16.0 | 24.9 | 13.4 | 6802 | 2955 |
| NeuralSparse | GraphSAINT | 36.8 | 14.6 | 22.9 | 9.1 | 7713 | 1943 |
TABLE II: Training results
| Sparsification Method |
Sampling Method |
Time per Epoch (s) |
Number of Epochs |
ROC AUC (%) |
EDP (W min²) |
|---|---|---|---|---|---|
| None | Uniformly Random | 18.1 | 49 | 66.56 | 3605.21 |
| None | GraphSAGE | 14.7 | 29 | 68.27 | 1726.45 |
| None | GraphSAINT | 1.4 | 45 | 72.23 | 48.29 |
| DropEdge | Uniformly Random | 36.4 | 17 | 64.15 | 6413.80 |
| DropEdge | GraphSAGE | 28.2 | 38 | 74.76 | 19553.45 |
| DropEdge | GraphSAINT | 1.8 | 31 | 70.53 | 39.01 |
| NeuralSparse | Uniformly Random | 106.2 | 25 | 76.46 | 81455.40 |
| NeuralSparse | GraphSAGE | 84.1 | 19 | 77.34 | 29008.15 |
| NeuralSparse | GraphSAINT | 57.1 | 37 | 79.17 | 46494.81 |