GraphBuff

Benchmarks

Every performance number on this site, with the graph it ran on, the hardware as far as it was recorded, the commit that published it and the command that reproduces it.

Graph kernels on the GPU

The same kernels on the CPU and on the GPU, with identical results (PageRank bit for bit).

Breadth-first search510 ms21 ms24×
Three-hop walk counts180 ms19 ms9.5×
Cypher two-hop count(*)154 ms31 ms5.0×
PageRank, 20 iterations1,293 ms557 ms2.3×
Graph
13 million relationships
Hardware
Radeon RX 5700 XT against 4 CPU threads. The CPU model is not recorded.
Published
2026-11-13, commit b3d47e1
Reproduce
cargo run --release -p graphbuff --features gpu --example gpubench

Kernels in query plans

The planner replaces multi-hop aggregates, filtered aggregates and Top-N with kernels over the adjacency export. CPU only.

Two-hop sum14,000 ms8 ms1750×
Filtered aggregate63 ms2.4 ms26×
ORDER BY … LIMIT 1068 ms3.3 ms21×
Graph
The csrbench graph: 100,000 nodes, 1.3 million relationships, in memory
Hardware
Not recorded.
Published
2026-11-13, commit 9e0902e
Reproduce
cargo run --release -p graphbuff --example csrbench

Graph algorithms

The algo.* procedures on the CPU.

PageRank (20 iterations)104 ms
Weakly connected components52 ms
Triangle count184 ms
Graph
The csrbench graph: 100,000 nodes, 1.3 million relationships, in memory
Hardware
Not recorded.
Published
2026-11-13, commit 9e3b10c
Reproduce
cargo run --release -p graphbuff --example csrbench

Adjacency export

Building the compressed adjacency that kernels and algorithms read. It is built once per committed state and shared.

From the in-memory store0.26 s
From the disk-resident store0.6 s
Size of the export32 MiB
Graph
The csrbench graph: 100,000 nodes, 1.3 million relationships
Hardware
Not recorded.
Published
2026-11-13, commit 2dfc778
Reproduce
cargo run --release -p graphbuff --example csrbench

Bulk import

graphbuff import writes store records directly, without Cypher or a redo log, compared with loading the same files through LOAD CSV and CREATE.

Load the graph~140 s~9 s16×
Graph
100,000 nodes and 1,000,000 relationships, with properties
Hardware
Not recorded.
Published
2026-11-13, commit 34e65f4
Reproduce
Not scripted yet: the engine has no benchmark program for this one.

How to read these

The numbers are the engine's own measurements, copied from its README. They are single runs on one developer machine, not an independent benchmark, and nothing here compares GraphBuff with another database. Where the hardware was not written down, this page says so. Keeping results in the repository is an open item in the engine's progress log.