odeinjul / DUCATI_SIGMOD

Accepted paper of SIGMOD 2023, DUCATI: A Dual-Cache Training System for Graph Neural Networks on Giant Graphs with the GPU

DUCATI

This repository contains implementation of DUCATI & SOTA and the overall training scripts. We put the underlying implementations of some APIs in a customized version of DGL (https://github.com/initzhang/dc_dgl.git).

Please follow these steps to prepare environment and datasets:

1. follow the instructions to install our customized DGL from source
2. install other dependent libraries in requirements.txt
3. prepare the datasets: download PA from OGB, download UK & UU & TW from GNNLab and WebGraph
4. pre-process the datasets using scripts in the preprocess directory

Then we can run the experiments under different settings as follows:

# verify dual cache allocation plan of DUCATI
CUDA_VISIBLE_DEVICES=0 python run_allocate.py --dataset [DS] --fanouts [FS] --fake-dim [FD] --total-budget [TB]

# verify iteration time of DUCATI using the allocation plan above 
CUDA_VISIBLE_DEVICES=0 python run_ducati.py --dataset [DS] --fanouts [FS] --fake-dim [FD] --adj-budget [AB] --nfeat-budget [NB]

# verify iteration time of SOTA
CUDA_VISIBLE_DEVICES=0 python run_sota.py --dataset [DS] --fanouts [FS] --fake-dim [FD] --nfeat-budget [NB]

---

The detailed cache configurations used in the paper:

While it is possible to determine the total budget inside scripts on the fly (run a normal training script without cache, print the torch.cuda.max_memory_allocated() after training, then substract the printed value from total GPU memory), we find that it is more flexible to expose total budget as a configurable knob during our experiments.