Benchmark NeuG: The Dual-Mode Graph Database

This tutorial demonstrates how to reproduce NeuG’s benchmark performance results. NeuG supports two execution modes:

• Embedded Mode: Run queries directly in your Python process with zero serialization overhead
• Service Mode: Connect to a NeuG server for concurrent transactional workloads

We benchmark both modes against industry-standard databases to showcase NeuG’s performance advantages.

Note: The LDBC SNB Interactive benchmark section covers only the complex read queries (IC1–IC14). It does not include write operations and transactional updates. A full LDBC SNB Interactive benchmark tutorial will be provided in a future release.

Dataset: LDBC SNB SF1

Both benchmarks use the same LDBC SNB SF1 dataset:

• Nodes: ~3 million (Person, Post, Comment, Tag, etc.)
• Edges: ~17 million (KNOWS, LIKES, HASTAG, etc.)
• Size: ~282MB compressed, ~1.1GB extracted

Download the Dataset

wget https://neug.oss-cn-hangzhou.aliyuncs.com/datasets/ldbc-snb-sf1-lsqb.tar.gz
tar -xzf ldbc-snb-sf1-lsqb.tar.gz

LSQB Benchmark (Embedded Mode)

LSQB  contains 9 complex subgraph matching queries that lean toward analytical workloads. This benchmark compares NeuG with LadybugDB in embedded mode.

Note on KNOWS Edges: The original LSQB benchmark assumes KNOWS relationships are bidirectional (i.e., if A knows B, then B also knows A). In our tests, we modified all queries involving KNOWS edges to use directed traversal (-[:KNOWS]->). This adjustment allows the same LDBC SNB SF1 dataset to be used for both SNB Interactive and LSQB benchmarks, since the KNOWS relationships in the original LDBC SNB data are unidirectional. This modification does not affect the fairness of evaluating graph database query optimization and execution capabilities.

Query Descriptions

Running the Benchmark

# Create virtual environment
python -m venv neug-env
source neug-env/bin/activate
 
# Install dependencies
pip install neug real_ladybug
 
# Run the benchmark
cd neug/examples/lsqb_benchmark
python run_benchmark.py --data-dir /path/to/ldbc-snb-sf1-lsqb

Use --force to overwrite an existing database.

Expected Results

On Apple Silicon Mac (M1/M2/M3), NeuG wins 8 out of 9 queries with just a single thread, even when comparing against LadybugDB’s best multi-threaded result:

NeuG excels on complex join-heavy queries, achieving dramatic speedups up to 287x on triangle patterns (Q3) and 91x on multi-hop filtering (Q2). LadybugDB’s multi-threading advantage shows on simpler traversal queries like Q6.

LDBC SNB Interactive Benchmark (Service Mode)

LDBC SNB Interactive  contains 14 complex read queries (IC1–IC14) covering multi-hop friend finding, shortest paths, and aggregation. This benchmark compares NeuG with Neo4j in service mode.

Query Descriptions

Running the Benchmark

# Start Neo4j server (optional, for comparison)
docker run -d --name neo4j \
  -p 7474:7474 -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/neo4j123 \
  neo4j:latest
 
# Create virtual environment
python -m venv neug-env
source neug-env/bin/activate
 
# Install dependencies
pip install neug neo4j
 
# Run the benchmark
cd neug/examples/ldbc_interactive_benchmark
python run_benchmark.py --data-dir /path/to/ldbc-snb-sf1-lsqb

Expected Results

On Apple Silicon Mac (M1/M2/M3), with 4 concurrent clients for 300 seconds:

Throughput Comparison

NeuG achieves 50.6x the throughput of Neo4j. On latency, NeuG’s P95 is just 20.6ms while Neo4j’s P95 reaches 1,728ms.

Per-Query Latency (P50 ms)

NeuG wins on 10 out of 14 queries, with dramatic speedups on complex multi-hop and aggregation queries.

Why NeuG is Faster

NeuG’s performance advantage comes from:

Embedded Mode Advantages

1. Graph-native query optimizer: GOpt  performs cost-based optimization with graph-specific cardinality estimation.

2. Columnar vectorized execution: Efficient memory management and cache-friendly traversals.

3. Zero serialization overhead: Executes in-process with direct memory access.

Service Mode Advantages

1. MVCC concurrency control: Efficient handling of concurrent transactions without blocking reads.

2. Battle-tested engine: NeuG is built on GraphScope Flex , the engine that set the LDBC SNB Interactive world record  at 80,000+ QPS.

Reproducibility

All performance results are independently reproducible:

• Dataset: https://neug.oss-cn-hangzhou.aliyuncs.com/datasets/ldbc-snb-sf1-lsqb.tar.gz 
• LSQB Scripts: neug/examples/lsqb_benchmark/
• Interactive Scripts: neug/examples/ldbc_interactive_benchmark/
• Tutorial: neug/doc/source/tutorials/benchmark-neug-dual-mode.md

If you encounter any issues reproducing these results, please report them on GitHub Issues .