Edge Partitioning Algorithms for Large Graphs

These algorithms are implemented by Qin Liu during his study at CUHK.

In this repo, we implement several edge partitioning algorithms and compute their replication factors for comparison:

Random partitioning (random)
Degree-based hashing (DBH): a paper on NIPS'14
A method based on Hilber space-filling curve (HSFC): this one is inspired by Frank McSherry's post
Our algorithms described in our paper Graph Edge Partitioning via Neighborhood Heuristic (published in KDD'17)
- Neighbor expansion (NE): described in Section 3 of our paper
- Streaming neighbor expansion (SNE): described in Appendix B of our paper

Compilation and Usage

We tested our program on Ubuntu 14.04/16.04, and it requires the following packages: cmake, glog, gflags, boost:

sudo apt-get install libgoogle-glog-dev libgflags-dev libboost-all-dev

Compilation:

git clone https://github.com/ansrlab/edgepart.git
cd edgepart
mkdir release && cd release
cmake ..
make -j8

Usage:

$ ./main --help
main: -filename <path to the input graph> [-filetype <edgelist|adjlist>] [-p <number of partitions>] [-memsize <memory budget in MB>]

  Flags from /home/qliu/workspace/edgepart/src/main.cpp:
    -filename (the file name of the input graph) type: string default: ""
    -filetype (the type of input file (supports 'edgelist' and 'adjlist'))
      type: string default: "edgelist"
    -inmem (in-memory mode) type: bool default: false
    -memsize (memory size in megabytes) type: uint64 default: 4096
    -method (partition method: ne, sne, random, and dbh) type: string
      default: "sne"
    -p (number of parititions) type: int32 default: 10
    -sample_ratio (the sample size divided by num_vertices) type: double
      default: 2

Example. Partition the Orkut graph into 30 parts using our NE algorithm:

$ ./main -p 30 -method ne -filename /path/to/com-orkut.ungraph.txt

Example. Partition the LiveJournal graph into 30 parts using our SNE algorithm (CacheSize = 2|V|, see our paper for detailed description):

$ ./main -p 30 -method sne -filename /path/to/com-lj.ungraph.txt -sample_ratio 2

Evaluation

The experiments are conducted on a PC with 16 GB RAM. MLE means "memory limit exceeded".

Algorithms that are listed but not contained in this repo:

METIS
Sheep: published on VLDB'15
Algorithms that has been integrated in to PowerGraph
- Oblivious
- High-Degree (are) Replicated First (HDRF): published on CIKM'15

Algorithm	wiki-Vote	email-Enron	web-Google	com-LiveJournal	com-Orkut	twitter-2010	com-Friendster	uk-union
METIS	5.25	2.40	1.05	2.13	MLE	MLE	MLE	MLE
NE	2.35	1.34	1.12	1.55	2.48	1.88	1.98	1.04
Random	9.28	5.10	6.54	8.27	19.48	11.68	11.84	15.99
DBH	5.43	3.32	4.09	5.18	11.97	3.67	6.88	5.14
Oblivious	3.85	2.30	2.28	3.43	6.94	8.60	8.82	2.03
HDRF	3.90	2.12	2.18	3.33	7.27	7.90	8.87	1.62
Sheep	4.20	1.78	1.71	3.33	7.94	2.34	4.45	1.29
SNE	3.05	1.44	1.17	1.88	4.49	2.83	3.00	1.65
HSFC	3.80	3.75	2.66	3.78	6.31	5.82	4.80	1.96