ming-make / gnnear-pact-2022

GNNear: Accelerating Full-Batch Training of Graph NeuralNetworks with Near-Memory Processing

Artifact of 'GNNear' paper for PACT 2022

This is the artifact of paper "GNNear: Accelerating Full-Batch Training of Graph Neural Networks with Near-Memory Processing", which can reimplement the experiments in Sec.6.

Requirements

We provide a docker image to setup the environment. Please run ./enter_docker_env.sh to enter the environment. You can also build the docker image by running ./build_docker_gpu.sh in the docker_scripts folder

Usage

DGL Training Profile (Optional)

You can modify and run run_dgl_train_profile.sh to conduct these CPU/GPU training experiments.

Alternatively, you can run train_epoch.py and set arguments to decide the configurations you want to run:

python train_epoch.py [--dataset_root_dir DATASET_ROOT_DIR] [--device DEVICE] 
                      [--layers LAYERS] [--dataset_name DATASET_NAME]
                      [--model_name MODEL_NAME] [--hidden_dim HIDDEN_DIM]

• --dataset_root_dir DATASET_ROOT_DIR: Set the root directory to save datasets used in our experiments. The default directory is ./dataset. Our code will automatically download missing datasets.

• --device DEVICE: Choose the device (cpu or gpu) you want to run training task on. The default device is cpu.

• --layers LAYERS: Set the layer number of the model. The default layer number is 2.

• --dataset_name DATASET_NAME: Choose the dataset you want to use. The default dataset is nell.

• --model_name MODEL_NAME: Choose the GNN model you want to train. The default model is GCN.

• --hidden_dim HIDDEN_DIM: Set the input/output dim of hidden layers. The default dim is 256.

This step is not necessary since we will use the previously profiled results for comparison to avoid performance number mismatching caused by hardware differences and other unstable factors.

DRAMsim3 Simulation (Optional)

To validate our architectue design, we implement a fine-grained simulator based on DRAMsim3 and cnpy. This simulator can simulate one channel's DIMMs' behaviour on different graph datasets.

You can run run_dramsim3_simulation.sh to start simulation. The results will be saved in results in the main directory. The experiments will take a long time to finish. This step is optional, you can skip it if you do not want to wait.

You can also run the steps manually:

• cd dramsim3_simulation/ext/cnpy, then follow the README.md to generate dynamic library of cnpy. You can directly run compile.sh in this folder.

• cd graph_partition, use the scripts to split graph data and save them in .npz format. Currently, our simulator injects the meta-data of Amazon and Reddit datasets, you can also add other datasets in dramsim3_simulation/src/graph_settings.h.

• Set the channel id and dataset you want to run by setting macros in dramsim3_simulation/src/graph_settings.h.

• Compile the simulator using compile.sh in dramsim3_simulation, then run binary dramsim3main to start simulate.

You can see the results in dramsim3_simulation/results. Each DIMM will report in DRAMsim3's manner.

Coarse-Grained Simulation

We provide Python scripts to fastly simulate DRAMSim3's behaviour in a corse-grained manner. You can run GNNear_experiments.py in simulation and set arguments to decide which experiment you want to run:

python GNNear_experiments.py [--dataset_root_dir DATASET_ROOT_DIR]
                             [--throughput] [--breakdown] [--ieo] [--shard]
                             [--window] [--ratio] [--rank]

• --dataset_root_dir DATASET_ROOT_DIR: Set the root directory to save datasets used in our experiments. The default directory is ~/datasets/Graph/. Our code will automatically download missing datasets.

• --throughput: If this option is on, training throughput comparison experiment in Sec.6.2 will be conducted, and the results will be saved in ./results/throughput/.

• --breakdown: If this option is on, speedup breakdown over CPU experiment in Sec.6.3 will be conducted, and the results will be saved in ./results/breakdown/.

• --ieo: If this option is on, effect of IEO experiment in Sec.6.3 will be conducted, and the results will be saved in ./results/ieo/.

• --shard: If this option is on, shard size exploration experiment in Sec.6.4 will be conducted, and the results will be saved in ./results/shard/.

• --window: If this option is on, window size exploration experiment will be conducted, and the results will be saved in ./results/window/

• --ratio: If this option is on, duplication ratio exploration experiment will be conducted, and the results will be saved in ./results/ratio/

• --rank: If this option is on, ranks per DIMM experiment will be conducted, and the results will be saved in ./results/rank/

Alternatively, you can also modify and run run_coarse_grained_simulation.sh to directly conduct all of these experiments.

The experiments will take about 10 hours to finish. You can run the experiments in a tmux environment (typetmux to enter the shell and run tmux a to re-enter the shell. The processes running in a tmux shell will not be terminated even if you close the remote ssh).

Results Validation

After the experiments finish, you can compare the results figures with the papers'. results/throughput should match Figure-17. results/breakdown should match Figure-19-a, results/ieo should match Figure-19-b. results/{rank,ratio,shard,window} should match Figure-21.