Skip to content

Getting Started

Jump to bottom
Abdullah edited this page May 23, 2026 · 1 revision

Getting Started

Build GraphBrew and run your first benchmark.

Requirements

Component Minimum Notes
OS Linux (Ubuntu 20.04+) or macOS tested on Ubuntu 22.04 LTS
Compiler GCC 7+ C++17 required
RAM 8 GB 16 GB+ for large graphs, 64 GB+ for the paper's full suite
Disk 50 GB benchmark graphs
Python 3.8+ (optional) analysis scripts use stdlib only

Optional: numpy, matplotlib, pandas for plotting and analysis.

Install dependencies

# Ubuntu / Debian
sudo apt-get install -y build-essential g++ libboost-all-dev \
    libnuma-dev google-perftools python3

# Fedora / RHEL
sudo dnf install -y gcc-c++ make boost-devel numactl-devel gperftools python3

# macOS
xcode-select --install
brew install gcc boost google-perftools

Verify with python3 scripts/graphbrew_experiment.py --check-deps.

Build

git clone https://github.com/UVA-LavaLab/GraphBrew.git
cd GraphBrew
make all        # production binaries → bench/bin/
make all-sim    # cache-sim binaries → bench/bin_sim/

Build flags:

Flag Effect
RABBIT_ENABLE=0 drop the Boost-backed Rabbit variant. The CSR variant stays enabled.
DEBUG=1 debug symbols
SANITIZE=1 address sanitizer

Boost 1.58 is only needed for -o 8:boost; the default -o 8 (CSR Rabbit) needs no Boost. If you do need it: python3 scripts/graphbrew_experiment.py --install-boost downloads and installs it to /opt/boost_1_58_0.

Verify

ls bench/bin/   # expect: bc bfs cc cc_sv converter pr pr_spmv sssp tc tc_p
ls bench/bin_sim/   # expect: bc bfs cc cc_sv pr pr_spmv sssp tc

# Smoke test on the included tiny graph
./bench/bin/pr -f scripts/test/graphs/tiny/tiny.el -s -n 3

You should see Read Time, Build Time, three Trial Time lines, and Average Time.

First benchmark

PageRank on the tiny graph with three reordering choices:

GRAPH=scripts/test/graphs/tiny/tiny.el

# No reordering (baseline)
./bench/bin/pr -f $GRAPH -s -o 0 -n 3

# Hub-clustered (degree-based; cheap, decent on power-law graphs)
./bench/bin/pr -f $GRAPH -s -o 7 -n 3

# GraphBrewOrder (Leiden + per-community ordering)
./bench/bin/pr -f $GRAPH -s -o 12 -n 3

-o N selects the reordering algorithm (see Reordering-Algorithms). -s symmetrises a directed input. -n N runs N trials.

Run a real graph

# Download a small social network from SNAP
wget https://snap.stanford.edu/data/ego-Facebook.txt.gz
gunzip ego-Facebook.txt.gz
mv ego-Facebook.txt facebook.el

./bench/bin/pr -f facebook.el -s -o 12:hrab -n 5

12:hrab is the HRAB variant — Leiden + Rabbit-on-supergraph. Other variants: 12:leiden, 12:rabbit, 12:tqr, 12:hcache, 12:rcm, 12:hubcluster, 12:streaming. See GraphBrewOrder.

Common flags

Flag Purpose
-f <file> input graph
-s symmetrise directed input
-o <algo> reordering algorithm (or -o 12:variant)
-n <N> number of trials
-r <vertex> root for BFS / SSSP
-i <iters> iterations per trial (PR)

Full list: Command-Line-Reference.

Automated pipeline

For batch experiments (download → build → benchmark → analyse):

# Quick test (~10 min): small graphs only
python3 scripts/graphbrew_experiment.py --target-graphs 30 --size small

# Standard (~1 h): medium graphs
python3 scripts/graphbrew_experiment.py --target-graphs 80 --size medium

# Full (~4-8 h): all sizes, 150 graphs per bucket
python3 scripts/graphbrew_experiment.py --target-graphs 150

# Preview only — print commands without running
python3 scripts/graphbrew_experiment.py --target-graphs 150 --dry-run

--target-graphs N is shorthand for --full --catalog-size N --auto --all-variants. See Benchmark-Suite for size buckets.

Reproducing the VLDB 2026 paper

# Preview: 2 small graphs, 1 trial — sanity check (~5 min)
python3 scripts/experiments/vldb/runner.py --all --preview

# 6-graph local set (cit-Patents → com-Orkut, fits 64 GB)
python3 scripts/experiments/vldb/runner.py --all --local

# 11-graph 64 GB-RAM set (adds web/mesh/synthetic, all auto-downloadable)
python3 scripts/experiments/vldb/runner.py --all --64gb

# Full set (includes twitter7 + webbase-2001, needs 256 GB+ RAM)
python3 scripts/experiments/vldb/runner.py --all

See VLDB-Experiments for the experiment-by-experiment guide.

Common build issues

Error Fix
fatal error: boost/range/algorithm.hpp sudo apt-get install libboost-all-dev
-fopenmp not supported sudo apt-get install libomp-dev
g++ too old (< 7) install g++-11, export CXX=g++-11, rebuild
OOM during build drop parallelism: make -j2 all

More in Troubleshooting.

Next steps

Clone this wiki locally