General GNN Accelerator.

Installation Guide

Anaconda Source Changing

# use Python3
wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh
bash miniconda.sh -b -p $HOME/usr/miniconda
[edit ~/.bashrc]
export PATH=$HOME/usr/miniconda/bin:$PATH

# use SJTU conda mirror
conda config --remove-key channels
conda config --add channels https://mirror.sjtu.edu.cn/anaconda/pkgs/free
conda config --add channels https://mirror.sjtu.edu.cn/anaconda/pkgs/main
conda config --add channels https://mirror.sjtu.edu.cn/anaconda/pkgs/mro
conda config --add channels https://mirror.sjtu.edu.cn/anaconda/pkgs/msys2
conda config --add channels https://mirror.sjtu.edu.cn/anaconda/cloud/pytorch/
conda config --add channels https://mirror.sjtu.edu.cn/anaconda/cloud/conda-forge/ 
# show the channel urls
conda config --set show_channel_urls yes

# create a virtual environment
conda create -n sga python=3.8
conda activate sga
# remove a virtual environment
# conda env remove -n sga

Core Packages

PyTorch, PyTorch Geo (PyG), Open Graph Benchmark (OGB)

conda install rdflib pyg -c pyg # pyg
pip install pytorch torchvision torchaudio # pytorch
pip install --upgrade ogb # ogb

Misc.

git config --global core.autocrlf input # git push setting

conda install graphviz python-graphviz # graphviz

# Other common packages (optional)
conda install xlutils matplotlib pandas requests networkx scipy numba

Run Simulation

1. We provide the scripts build.sh and clean.sh for convenience.

2. Then go to the folder of cppmain/evaluation and run execution or partition with proper arguments.

3. You can run make in cppmain/evaluation to rebuild.

See examples in cppmain/scripts.

Multi-Tasking

We recommend tmux to launch and manage processes in background. We do not use the nohup command anymore for the difficulty in scaling up the experiments.

chmod +x *.sh

Memory Leakage Analysis

We recommend valgrind to check the memory usage of the simulation. Before starting up, you may put the following line into the CMakeLists.txt and rebuild to expose the information to valgrind.

add_compile_options(-g)

Graph Dataset (Directed)

Name	Vertex#	Edge#	Feature#	Class#
Cora (CR)	2,708	10,556	1,433	7
CiteSeer (CS)	3,327	9,104	3,703	6
PubMed (PM)	19,717	88,648	500	3
Flickr (FL)	89,250	899,756	500	7
Arxiv (AX)	169,343	1,166,243	128	40
Reddit (RD)	232,965	23,213,838	602	41
Yelp (YP)	716,847	13,954,819	300	-
Product (PD)	2,449,029	61,859,140	100	47

Legacy

Name	Vertex#	Edge#
cit-Patents (CP)	3,774,768	16,518,948
soc-LiveJournal1 (SL)	4,847,571	43,369,619
hollywood (HW)	1,139,905	57,515,616
europe_osm (EO)	50,912,018	54,054,660
soc-twitter-2010 (ST)	21,297,772	265,025,809

Change Log

What is in the version 9:

• Tile coalesce
• Thread <= Stream

What is in the version 8:

• Code refactored in hardware perspective.
  • Class == Module
• Online tiling via TilingUnit (TU).
  • Supporting CSR, CSC formats
• Compiler from DGL to IR.
• ISCA 2022 submission.

What is in the version 7:

• Multiple graph reordering/clustering algorithms.
  • basic reordering (Sort, Random).
  • hub-based sorting (HubSort, HubCluster).
  • degree-based grouping (DBG).
  • minhash grouping (TVC).

What is in the version 6:

• New controller for the multi-stream execution.
• MICRO 2021 submission.

What is in the version 5:

• Instruction driven Architecture.
• Tile-based multi-stream execution.
• Degree-sorting-based vertex reordering.
• High-performance C++ simulator.

What is in the version 4:

• Hint-guided execution with hilbert-curve tiling.
• Decoupling the hardware execution and software partition.
• Ramulator DRAM simulator integration.

What is in the version 3:

• Unified SA for both ApplyEdge & ApplyVertex
• Energy consumption in hardware execution.
• HPCA 2021 submission.

What is in the version 2:

• An extra SA for SAGA-NN execution.
• Hilbert space-filing curve for graph partition.

What is in the version 1:

• HyGCN performance simulation reproduction.
  • with pipelined SIMD and SA design