ldbc

LDBC Benchmark Results — Samyama v0.6.0

Samyama's query engine benchmarked against all four LDBC Council benchmark suites.

Test Environment: Mac Mini M4 (10-core, 16GB RAM), macOS Tahoe 26.2, Rust 1.85 release build

Date: 2026-03-07

Summary

Benchmark	Queries	Passed	Pass Rate	Dataset	Total Time
SNB Interactive	21 reads + 8 inserts + 8 deletes	21/21 reads	100%	SF1 (3.18M nodes, 17.26M edges)	111.9s
SNB Business Intelligence	20	16/16 run (BI-17+ timeout)	100% of run	SF1 (same dataset)	~52s (16 queries)
Graphalytics	28 (6 algos x 5 datasets)	28/28	100%	XS (2) + S-size (3) datasets	<1ms (XS), 0.1–167s (S)
FinBench	40 (12 CR + 6 SR + 3 RW + 19 W)	40/40	100%	Synthetic (7.7K nodes, 42.2K edges)	665ms

Overall Coverage

• 4 LDBC benchmark suites implemented
• 82 unique query/algorithm implementations across all suites (including 8 deletes)
• Graphalytics 28/28 passing — XS 12/12 + S-size 16/16 (PageRank exact iterations, dangling redistribution, directed LCC)
• S-size datasets validated: wiki-Talk (2.4M V), cit-Patents (3.8M V), datagen-7_5-fb (633K V, 68M E)
• WITH projection barrier implemented for BI query support
• GPU acceleration (Enterprise) for PageRank, LCC, CDLP, WCC, BFS

Quick Start

# SNB Interactive (21 read queries, SF1 dataset required)
cargo bench --release --bench ldbc_benchmark -- --runs 3

# SNB Interactive with update + delete operations
cargo bench --release --bench ldbc_benchmark -- --runs 3 --updates --deletes

# SNB Business Intelligence (20 analytical queries, 120s timeout per query)
cargo bench --release --bench ldbc_bi_benchmark -- --runs 3

# Graphalytics (6 algorithms, XS datasets)
bash scripts/download_graphalytics.sh
cargo bench --release --bench graphalytics_benchmark -- --all

# Graphalytics (S-size datasets)
bash scripts/download_graphalytics.sh --size S
cargo bench --release --bench graphalytics_benchmark -- --size S --all

# FinBench (21 queries, synthetic data auto-generated)
cargo bench --release --bench finbench_benchmark -- --runs 3

Detailed Reports

• SNB Interactive — Full Results
• SNB Business Intelligence — Full Results
• Graphalytics — Full Results
• FinBench — Full Results

Benchmark Architecture

All benchmarks share a common pattern:

1. Load dataset into in-memory GraphStore via EmbeddedClient
2. Warm up each query once (populates AST cache)
3. Benchmark each query N times, recording min/median/max latency
4. Report results in a formatted table

Data loading uses the shared ldbc_common module for CSV parsing and graph population. Queries use query_readonly() (read lock) for reads and query() (write lock) for updates.