LDBC Graphalytics ArcadeDB Platform Driver

Platform driver implementation for the LDBC Graphalytics benchmark using ArcadeDB.

Uses ArcadeDB in embedded mode with the Graph Analytical View (GAV) engine, which builds a CSR (Compressed Sparse Row) adjacency index for high-performance graph algorithm execution with zero GC pressure.

This repository contains three benchmark modes:

1. Official LDBC Graphalytics — standardized framework with per-algorithm isolation, validation, and reporting
2. Native multi-vendor comparison — load once, run all algorithms, compare ArcadeDB vs Kuzu vs DuckPGQ vs Memgraph vs Neo4j vs FalkorDB vs HugeGraph
3. LSQB (Labelled Subgraph Query Benchmark) — 9 subgraph pattern matching queries on the LDBC SNB social network, comparing ArcadeDB (Cypher) vs DuckDB (SQL) vs FalkorDB (Cypher) and others

Supported Algorithms

Prerequisites

• Java 21 or later (required for jdk.incubator.vector SIMD support). The published benchmark numbers use Eclipse Temurin 25 with -XX:+UseCompactObjectHeaders (GraalVM is not used)
• Maven 3.x
• Python 3.10+ (for Mode 2 and Mode 3 multi-vendor comparisons; see pyproject.toml for per-vendor install extras)

Build

mvn package -DskipTests

The build produces a self-contained distribution in graphalytics-1.3.0-arcadedb-0.1-SNAPSHOT/.

Dataset

Use the built-in dataset manager to browse and download datasets from the LDBC data repository:

# See all available datasets (40+ Graphalytics + 9 LSQB scale factors)
python3 datasets.py available

# Download the standard Graphalytics benchmark dataset (633K vertices, 34M edges, ~155 MB)
python3 datasets.py download datagen-7_5-fb

# Download the LSQB social network dataset (SF1, ~3.9M vertices, ~17.9M edges)
python3 datasets.py download lsqb-sf1

# Show downloaded datasets with size and vertex/edge counts
python3 datasets.py

Datasets are downloaded into the datasets/ directory (git-ignored). After downloading datagen-7_5-fb:

datasets/
  datagen-7_5-fb/
    datagen-7_5-fb.v              # vertex file (one ID per line)
    datagen-7_5-fb.e              # edge file (src dst weight, space-separated)
    datagen-7_5-fb.properties     # graph metadata
    datagen-7_5-fb-BFS/           # validation data per algorithm
    datagen-7_5-fb-WCC/
    ...

---

Weekly unattended run

python3 weekend.py --jvm-flags "-XX:+UseCompactObjectHeaders"

Runs every suite and vendor in sequence (ArcadeDB embedded Java benchmarks, the Graphalytics and LSQB multi-vendor suites, optionally Mode 1 with --mode1-dist) and writes a Markdown/JSON report to weekly-results/<timestamp>/. Each vendor runs in its own process group with a total-time and a no-output limit; a hung vendor gets SIGTERM, then SIGKILL, and the run continues. Loaded databases are kept under ~/.cache/ldbc-graph-bench and reused, so only the first run (or --reset, a new image, a changed dataset) pays for the loads. See CLAUDE.md for the details and python3 -m unittest discover -s tests for the tests.

Benchmarks and results

Every number is a warm number (first call discarded, median of 3 timed runs), measured one system at a time on AC power and validated against the official LDBC reference outputs (or the official expected counts for LSQB). A value marked ✗ failed that check and is not ranked; N/A = no implementation; timeout = the 5-minute limit per operation. Footnote symbols (§ ‡ ¶ *) are explained on the detailed pages. Machine: MacBook Pro 16" (2026), Apple M5 Pro, 48 GB RAM; ArcadeDB on Eclipse Temurin 25 with -XX:+UseCompactObjectHeaders, 12 GB heap for every JVM system, Docker Desktop with 32 GB.

How ArcadeDB itself changes from release to release (official framework, Mode 2 and LSQB): ArcadeDB-release-progress.md.

Graphalytics, datagen-7_5-fb (633,432 vertices, 34,185,747 edges)

Seconds, last column peak memory in GiB. Bold marks the fastest valid result per column. Systems, versions, how we measure, per-system notes and how to run each vendor: detailed page.

Graphalytics, graph500-22 (2,396,657 vertices, 64,155,735 edges)

Seconds; no SSSP (not defined for this dataset). ArcadeDB is 26.11.1-SNAPSHOT; bold marks the fastest valid result per column. Per-system notes, the harness problems found and how to reproduce: detailed page.

OOM = out of memory (Docker Desktop's 32 GB). † remembered original load, ‡ DuckPGQ BFS does not finish (ended after 20 min), § Memgraph load includes the reverse-edge step, ¶ ArangoDB container ran out of memory, ** live heap after GC (process peak not measured): see the detailed page.

LSQB SF1 (3,947,829 vertices, 17,882,623 edges)

Seconds; ArcadeDB embedded is shown with the Graph Analytical View (OLAP) and without it (OLTP); bold marks the fastest result per query. All counts match the official expected output. Queries, run commands and analysis: detailed page.

More

• docs/benchmark-graphalytics-official.md, docs/benchmark-graphalytics-multivendor.md, docs/benchmark-graphalytics-graph500-22.md, docs/benchmark-lsqb.md: one page per benchmark (how to run, method, full tables, notes).
• docs/vendor-notes.md: SurrealDB and Dgraph (excluded from default runs, see also SURREALDB.md) and FalkorDB.
• docs/architecture.md: file structure, the Graph Analytical View (CSR) engine, execution modes.
• ArcadeDB-release-progress.md: ArcadeDB across releases. results-*.md: raw results and history of earlier runs. CLAUDE.md: benchmark rules for the harness.

License

Apache License, Version 2.0