Benchmarks
Every benchmark below is backed by reproducible benchmark scripts and raw result artifacts in benchmarks/results/ on BrickCoreStorage (Purple8 Cortex) with BrickWAL and CortexAutoConfig.
Canonical Headline Numbers (v0.72.0)
| Benchmark / Workload | Measured Result | Environment / Notes | Reference Artifact |
|---|---|---|---|
| Single-Node 20M-Node / 30M-Edge Ceiling (50M Objects) | 3,453 nodes/s · 9,291 edges/s (<1.1% drift) · 52.0 µs P50 (1.26 ms P99) point lookup · 0.38 ms P50 (0.87 ms P99) 2-hop walk · 4,180.7 QPS (13.8 ms P99) | 16 vCPU, 64 GB RAM (~42.3 GB RSS, 25.35 GB NVMe) | results/fly_20m_ceiling_20261030.json |
| Power-Law Super-Hub 2-Hop Walk (6M-Edge Hub @ 20M) | 53.1 ms P50 · 58.8 ms P95 · 58.8 ms P99 (0 timeouts) | Guarded intermediate hub traversal across a 6,000,000-edge central hub | results/fly_20m_ceiling_20261030.json |
| Single-Node 10M-Node / 15M-Edge Ceiling (25M Objects) | 3,604.7 nodes/s · 9,449.9 edges/s · 47.5 µs P50 lookup · 0.33 ms P50 2-hop walk · 3,786.4 QPS | 16 vCPU, 32 GB RAM (23.65 GB RSS) | results/fly_10m_ceiling_20261030.json |
| 250-Storey 3-D Structural Mega-Structure (AEC) | 240.52 ms 3-D rigidity · 3.53 s critical-member bridges · 916.65 ms Cuthill-McKee ordering | 100,000 3-D joints, 478,840 structural members in 352.5 MB peak RSS | results/fly_10m_ceiling_20261030.json |
| Vector query P99 @ 100K nodes | 0.33 ms | In-process GraphEngine (BrickCoreStorage) | results/BENCHMARK_SUMMARY_20260707.md |
| Batch ingest throughput | 40,551 docs/s (40K/s) | BrickCoreStorage + BrickWAL batch seal | results/BENCHMARK_SUMMARY_20260707.md |
| VectorDBBench 1M × 768d | 916 QPS, 0.946 recall@10 | Standard VectorDBBench 1M × 768d suite | results/BASELINE.md |
| HTTP P95 @ 1024d (full security stack) | 23 ms | JWT/API-key auth, BrickWAL, AES-256-GCM KMS encryption, and RBAC all enabled | results/BENCHMARK_SUMMARY_20260707.md |
| 16-vCPU / 32 GB RAM Enterprise Suite | 41,284 nodes/s burst ingest · 1,096 QPS concurrent hybrid search (P99 = 30.85 ms @ 32 workers) · 100% accuracy | 16 vCPU, 32 GB RAM; 100% accuracy across graph algorithms, GraphRAG, Journey SLA/HITL, SOC detector, and Quantum QUBO solvers | results/FLY_PERF16X_BENCH_20261025.md |
| Largest single-node corpus (8.84M MS MARCO) | ~9 GB peak RSS (bounded), -3,626 MB returned to OS after ingest, ~5.4 GB steady state, 0 swap | 8.84M docs (24 GB corpus) on a 24 GB machine via CortexAutoConfig brick seal-and-evict | results/industry_msmarco_8841823_*.json |
| 30-minute sustained soak test | 1.18M requests, 500 concurrent users, 0% error rate | 30-min continuous mixed workload | results/soak_20260705T144612/soak_summary.md |
| Durable REST ingest (1M nodes) | 3,270 nodes/s sustained, 0 errors, p99 = 98 ms @ 100 users | AWS c6i.4xlarge, Linux | results/EC2_BENCH_20260829.md |
Durable add_node (1 → 64 writers) | 2,393 → 17,204 req/s, p99 ≤ 5.5 ms | Linux, real fdatasync | BENCHMARK_METHODOLOGY.md |
| WAL crash recovery (SIGKILL) | 10,000 / 10,000 nodes survive mid-ingest SIGKILL + cold-start 0.33 s | BrickWAL replay verification | cortex_resilience_bench.py |
Enterprise 16-vCPU / 32 GB RAM Benchmark (FLY_PERF16X_BENCH_20261025)
On a dedicated 16-vCPU / 32 GB RAM production instance, Purple8 Hyper Graph v0.72.0 was validated end-to-end across throughput, concurrency, and functional accuracy:
- Burst Ingestion: 41,284 nodes/s with
BrickCoreStorageandBrickWAL. - Concurrent Hybrid Search: 1,096 QPS at 32 concurrent workers with
P99 = 30.85 ms(combining HNSW vector search, BM25 full-text, and graph traversal). - 100% Subsystem Accuracy:
- Graph Algorithms: 100% accuracy across PageRank, Louvain community detection, Brandes betweenness centrality, Dijkstra shortest path, link prediction, and Tarjan biconnected components/bridges.
- GraphRAG & Hybrid RAG: 100% citation and multi-hop retrieval accuracy.
- Journey Engine: 100% accuracy on stage transitions, SLA breach detection, and HITL gate enforcement.
- SOC Detector: 100% precision/recall across Layer 1 threshold rules, Layer 2 EWMA z-score anomalies, and Layer 3 MCP agent behavioural profiling.
- Quantum QUBO Solvers: 100% optimal/reference convergence across simulated annealing, tabu, greedy, and hybrid solvers.
Bounded Memory Envelope (8.84M MS MARCO Proof)
BrickCoreStorage (Purple8 Cortex) paired with CortexAutoConfig prevents RSS from scaling monotonically with total corpus size $N$. During ingestion of 8.84 million MS MARCO passages (24 GB on disk) on a 24 GB machine:
- Sawtooth RSS across 90 brick seal-and-evict cycles: Peak RSS remained bounded at ~9 GB throughout the entire run.
- Memory reclamation:
-3,626 MBof working-set memory was returned to the OS immediately after ingestion completed. - Query steady state: ~5.4 GB RSS with zero swap usage.
In-Process Micro-Benchmarks (Suites C–J)
| Suite | Subsystem | Measured Result |
|---|---|---|
| C | Write path (BrickCoreStorage) | add_node 0.045 ms p50 |
| D | Document point lookup | 0.07 ms p50 |
| E | Journey Engine | 0.23 ms stage advance |
| F | AES-256-GCM envelope encryption | ~0% single-node overhead, +1.2% batch overhead |
| G | Multi-tenancy isolation | 0 cross-tenant leaks across concurrent readers/writers |
| H | Graph analytics | Dijkstra shortest path 0.79 ms p50 @ 10K nodes |
| I | MCP server round-trip (125 tools, 16 namespaces) | graph.get_node 0.007 ms p50 |
| J | Rust core (rust_core/) | SLA scan 1.58× speedup |