hellgraph-bench

Reproducible benchmark harnesses and receipts for HellGraph vs. the graph-DB / graph-analytics field. Rule of the house: measure it or don't claim it — every number here is backed by a run with the VMs torn down afterward.

Layout

Path	What's in it
bench/	HellGraph benchmark runners (LDBC Graphalytics 6-kernel + scale runs).
neo4j/	Neo4j GDS comparison setup (startup.sh).
results/	Signed receipts. RECEIPT-NEO4J.md = GDS 6-kernel head-to-head.
aws/	The cuGraph "burial" kit — see below.
scripts/	Helpers.

Receipts so far (measured)

• LDBC Graphalytics, 6 kernels vs Neo4j GDS — HellGraph wins all 6 (see results/RECEIPT-NEO4J.md).
• 68.7B-edge PageRank on 61 GCP nodes, ~$30, ~4.5 min.
• ~50–70 GTEPS at 127 nodes — past a DGX-2's 38, on graphs a DGX can't hold.
• Kuzu ~50×; Python graph stacks 4–116×; Neptune walled out (32B ceiling, ~9,000× the price).

The cuGraph run (aws/) — the one still open

cuGraph is GPU-native (memory-bandwidth-bound). The play is to out-rent raw throughput on CPU aggregate bandwidth, and to run our own hg_gpu (wgpu/WGSL) kernel on NVIDIA and AMD silicon cuGraph can't touch. aws/BURIAL-PLAN.md has the full strategy.

Config is entirely via environment — no in-file editing. Set once, then run:

export REGION=us-east-1 AZ=us-east-1a SUBNET=subnet-xxxx KEY=my-keypair
export S3=s3://my-bucket/cugraph-race CPU_AMI=ami-xxx GPU_AMI=ami-yyy IAM_PROFILE=hg-s3
./aws/launch-aws.sh setup      # placement group + EFA security group + S3 upload
./aws/launch-aws.sh cugraph    # p4d/p5 box, RAPIDS 6-kernel, GTEPS -> S3
./aws/launch-aws.sh cluster    # coordinator + N hpc7a EFA workers, dist_p2p
./aws/launch-aws.sh collect    # pull both scorecards
./aws/launch-aws.sh teardown   # TERMINATE EVERYTHING — always run this

The launch script aborts before any aws call if a required variable is unset, so a half-configured run cannot spend money. Always finish with teardown and confirm the instance list prints empty.