SALIENT is a distributed multi-GPU training and inference system for graph neural networks (GNNs), targeting at massive graphs. It effectively mitigates (and with sufficient memory bandwidth, nearly eliminates) the bottleneck of a PyTorch workflow on mini-batch generation and transfer, rendering the overall training/inference time dominated by GPU computation. The system is described in the paper Accelerating Training and Inference of Graph Neural Networks with Fast Sampling and Pipelining published at MLSys 2022. This repository contains the codes of SALIENT.

UPDATE: SALIENT has a successor SALIENT++, which distributes the node feature matrix across machines. SALIENT does not rely on graph partitioning tools if the main memory can store the entire matrix.

SALIENT can be run on a machine with one or multiple GPUs or on a GPU cluster.

• See this README for instructions to setup the system on a GPU machine.
• See this README for instructions to setup the system on the Satori cluster.

Pointers of example usage of SALIENT are given in these instructions.

SALIENT is designed to be fully compatible with PyG. In particular, defining a GNN architecture is done the same way as writing a usual PyG code.

• See this README for instructions to add a GNN architecture.
• See this README for instructions to add a dataset.

While this repository is being maintained, a frozen version that reproduces the key results in the paper can be found in a separate artifact repository.

Please cite this paper if you use the codes in your work:

@INPROCEEDINGS{Kaler2022,
  AUTHOR = {Tim Kaler and Nickolas Stathas and Anne Ouyang and Alexandros-Stavros Iliopoulos and Tao B. Schardl and Charles E. Leiserson and Jie Chen},
  TITLE = {Accelerating Training and Inference of Graph Neural Networks with Fast Sampling and Pipelining},
  BOOKTITLE = {Proceedings of Machine Learning and Systems 4},
  YEAR = {2022},
}

SALIENT is developed by the xGraph team at the MIT-IBM Watson AI Lab.

This research was sponsored by MIT-IBM Watson AI Lab and in part by the United States Air Force Research Laboratory and the United States Air Force Artificial Intelligence Accelerator and was accomplished under Cooperative Agreement Number FA8750-19-2-1000. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the United States Air Force or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein.