Stress Test — 1M Elements in a Single Federated Session¶

Status: RTree GPU path + Cockpit UI + Smart Overnight Bake — DONE S186. Sandbox: scripts/sandbox_1M.db (1,061,736 elements, 29 real buildings) Builder: scripts/build_sandbox_1M.py Library: library/library.blend (120,471 meshes, 276 MB, shared all buildings)

S186 result: Background parallel Blender sessions (up to 4 instances) can resolve a full 1-million-element city in under 10 minutes at ~75MB total file size. Previously trended to 3.5 hours with save crashes. Terminal (48K elements) bakes in 36.6s. Hospital (63K) in 123s. Each building bakes in a fresh empty scene — no O(n) scene graph penalty. Linked mesh refs to library.blend keep files small. Smart Overnight auto-detects when offline bake is 5x faster and offers to switch — viewport stays fully interactive during the background bake. See docs/RTree.md for full architecture, benchmarks, and proof.

Goal¶

Load 1,000,000 BIM elements into Blender with interactive viewport performance. All geometry is real — extracted from 29 IFC buildings, tiled into a city layout.

The Loading Problem¶

The Library button (S174) creates one Blender object per element. At 1M elements: - 1M objects in Outliner (unusable) - .blend save writes 1M objects (huge file, minutes to save) - No camera-distance LOD (every element rendered equally) - Instance creation: ~800 objects/s → 20+ minutes to load

The Solution: RTree GPU + Stingy Mesh Loader¶

Earlier Approach — GN Mode (S175–S184, halted)¶

Geometry Nodes (GN) were explored as the instancing layer: one point cloud per discipline, Instance on Points picking meshes by hash_index. At small scale (7K meshes) this loaded in under 2 seconds. At city scale (500+ modifier trees) GN evaluation overhead reached 8 minutes — unviable. DLOD (distance-based LOD swaps) and progressive make_local() were implemented to mitigate, but the root cause — GN re-evaluating the full tree every frame — could not be fixed from Python. GN Mode was halted at S184.

See internal/GN_LINK_INVESTIGATION.md and internal/DLOD_SPEC.md for the full investigation archive.

Current Solution — RTree GPU Path (S184+)¶

The RTree Query Engine resolved the speed problem by eliminating mesh loading entirely from the default viewport:

DB (1M elements, SQLite R-tree index)
    ↓  O(log n) spatial query
RTree GPU       — 1M wireframes, 13s load, instant orbit, zero mesh in RAM
    ↓  drill-down (click building → fly)
Stingy Loader   — exact IFC geometry on demand, <1s per MESH press
    ↓  SHRED when done
Smart Bake      — background Blender, 36s/48K, 4-core parallel, 1M city <10 min

Zero Blender objects in the default view — GPU line batches drawn directly
On-demand mesh via Stingy Loader: viewport-centre query, pre-warmed library meshes linked from library.blend, one named collection per MESH press
LOAD / SHRED is non-destructive — load geometry to inspect, shred to clean up
Smart Overnight Bake (S186): background subprocess, fresh scene per building, no O(n) scene-graph penalty. Linked mesh refs keep files at ~2MB per building

See docs/RTree.md for the full architecture, benchmarks, and proof.

City Layout¶

Built by scripts/build_sandbox_1M.py from 29 real extracted buildings:

CBD strip (13 buildings): - Hospital ×2 (63,917 each), Terminal (48,428), LTU_AHouse (4,785) - HospitalGarage ×2, Clinic (16,480), 6 office buildings - Total: 221,386 elements

Suburb rows (51 rows × 18 house types): - Duplex, SampleHouse, BimWhale variants, Schependomlaan, etc. - 16,544 elements per row - Total: 843,744 elements

Grand total: 1,065,130 elements

Measured Performance¶

RTree GPU Path (current — S184+)¶

Metric	Value
Wireframe load (1M elements)	13s
Orbit / pan	Instant (60 FPS)
MESH press (viewport-centre)	<1s
Terminal bake (48K elements)	36s
Hospital bake (63K elements)	123s
Full city bake (1M, 4 cores)	<10 min
Blender objects in default view	0

GN Mode (historical — S175–S184, halted)¶

Metric	GN link=True	Per-element link=False
Mesh load	1.74s	37.05s
Instance/GN build	0.05s	57.47s
Total load	2.19s	64.08s
Outliner items	6	48,428
Viewport	Frozen (8-min GN eval overhead)	60 FPS

GN was fast to load but froze on evaluation at scale. RTree eliminated the problem by not loading meshes at all — wireframes from the spatial index, mesh on demand.

References¶

RTree.md — primary viewer architecture, benchmarks, proof
DLOD Spec — distance LOD handler design (GN era, archived)
Full Loader 2 SRS — master loader spec
GN Link Investigation — link=True vs link=False analysis (archived)
scripts/build_sandbox_1M.py — city builder
scripts/pipeline_library.sh — extraction pipeline
MANIFESTO.md §The Backend — architecture philosophy: OS-level parallelism, compile-once, linked .blend