About

DistMILE: A Distributed Multi-Level Framework for Scalable Graph Embedding Yuntian He, Saket Gurukar, Pouya Kousha, Hari Subramoni, Dhabaleswar K. Panda, Srinivasan Parthasarathy Published in HiPC 2021

Abstract: DistMILE is a Distributed MultI-Level Embedding framework, which leverages a novel shared-memory parallel algorithm for graph coarsening and a distributed training paradigm for embedding refinement. With the advantage of high-performance computing techniques, DistMILE can smoothly scale different base embedding methods over large networks.

Link: https://par.nsf.gov/servlets/purl/10326130

Citation:

@inproceedings{he2021distmile, title={Distmile: a distributed multi-level framework for scalable graph embedding}, author={He, Yuntian and Gurukar, Saket and Kousha, Pouya and Subramoni, Hari and Panda, Dhabaleswar K and Parthasarathy, Srinivasan}, booktitle={2021 IEEE 28th International Conference on High Performance Computing, Data, and Analytics (HiPC)}, pages={282--291}, year={2021}, organization={IEEE} }

Required packages

• Horovod
• TensorFlow
• MVAPICH2-GDR
• mpi4py
• numpy
• scipy
• scikit-learn
• networkx
• gensim (For DeepWalk)
• theano (For NetMF)

How to Run

Run the main program using the below command:

mpirun_rsh -np ${NUM_GPU} ${GPUS} ${MPI_ENV} python main.py --data ${DATA} --basic-embed ${EMBED} --batch-size ${BATCH_SIZE} --coarsen-m ${COARSEN_M} --coarse-threshold ${COARSE_THRESHOLD} --workers ${NUM_THREADS} --num-threads ${NUM_THREADS} --coarse-parallel

Arguments and variables:

• -np ${NUM_GPU}: Number of machines for use.
• ${GPUS}: Hostnames of the machines.
• ${MPI_ENV}: Environmental variables for MPI. Please use "MV2_USE_CUDA=1 MV2_SUPPORT_DL=1 MV2_ENABLE_AFFINITY=0 MV2_HOMOGENEOUS_CLUSTER=1".
• --data: Dataset, located in dataset/{$DATA}.
• --basic-embed: Base embedding method, located in base_embed_methods.
• --batch-size: Batch size for distributed learning. Default is 100000.
• --coarsen-m: Coarsen Depth.
• --coarse-threshold: Threshold for coarsening a graph in parallel. Denoted as $n_{c}$ in the paper. Default is 10000.
• --workers: # threads for base embedding.
• --num-threads: # threads for coarsening and refinement. All threads are used by default.
• --coarse-parallel: Leverage the parallel version of graph coarsening.
• --large: Only necessary input is sent to TensorFlow model. Not used by default.