wpybtw / wholegraph

WholeGraph

WholeGraph is developed to help train large-scale Graph Neural Networks(GNN).

The software stack of WholeGraph shown below:

Based on CUDA and NVIDIA's hardware, a multi-GPU storage called WholeMemory is implemented. WholeMemory is used to support both graph structure and feature embedding storage. Based on the storage, WholeGraph Ops are built to support GNN applications, including sampling ops, embedding ops and some GNN layer ops On the other hand, sampled subgraph from WholeGraph can also be converted to support PyG or DGL layers. So, GNN applications can utilize ops from WholeGraph and other GNN frameworks.

How to Use

Environment

Hardware

It is suggested to use NVLink systems, like DGX-A100 or similar systems.

Software

It is recommended to use our Dockerfile

Compile

To compile WholeGraph, from source directory:

mkdir build
cd build
cmake ../
make -j

Or you can build release version by replacing cmake ../ by cmake -DCMAKE_BUILD_TYPE=Release .., which has slightly better performance.

Examples

Checkout GNN example for more details.