Large optical access networks, made inspectable.

Optravo provides an interactive environment for examining the structure, spatial organization, and relationships within optical access networks.

Guided demonstration of graph navigation, identifier search, element selection and inspection of supplied network metadata.

© OpenStreetMap contributors ODbL © OpenMapTiles

Visualization benchmark at 600,000 nodes.

Benchmark dataset
600,000 nodes, 1,203,000 linksBenchmark graph retained in one visualization scene. Links average ~2.0 per node, since the graph is a spanning tree plus ~3,000 cross-connects from ring and redundancy links rather than a pure tree. The 1,203,000 figure counts each of the 599,999 tree edges twice and the ~3,000 cross-connects once (2 × 599,999 + 3,000 ≈ 1,203,000).
CPU frame cost distribution
10.0 ms median, 16.8 ms at p9524.7 ms at p99 and 52.6 ms at p99.9
Frame budget compliance
94.9% within 16.67 ms (60 FPS)99.6% within 33.3 ms (30 FPS) across 10 runs
Primary benchmark system
HP OMEN 40L, 32 GB DDR4-3200, NVIDIA RTX 3060 12 GBWindows 11 Pro 26H2 Build 26300.9550, Chrome 154.0.8037.92, 3840 × 2160, production build

Benchmark CPU frame cost distribution at 600,000 nodes on the HP OMEN 40L

Median10.0 ms
95th percentile16.8 ms
99th percentile24.7 ms
99.9th percentile52.6 ms
Measure
CPU frame cost is the main-thread JavaScript render cost per frame, measured as a performance.now() delta across the render tick.
Scope
GPU completion is not captured, which is why GPU-class differences between systems appear compressed.
Timing
Values are in-browser main-thread timings rather than photon-to-eye end-to-end latency, and a 60 Hz panel floors perceived latency at 16.7 ms.
Runs
Ten 60-second runs (36,000 frames) were recorded with the complete 600,000-node scene resident on the HP OMEN 40L, with every display at 60 Hz and 3,600 frames per run.

Topology inspection for auditable network analysis.

Validate on import

No dataset reaches the map unchecked. Identifiers, coordinates and link endpoints are examined first, and a file that does not hold together is set aside rather than allowed to pass for a network it is not, so every view rests on data that has been verified.

Identify the dataset

The type of a dataset is read from its content rather than from its file name. Common export shapes are accepted, including files that describe nodes only, and what Optravo recognised is stated alongside the data, so the basis of a view is never left implicit.

Survey the whole network

Central offices, splitters, ONUs, intersections and every class of link share a single map. Dense areas gather into groups instead of dissolving into noise, and large datasets remain calm to navigate without the network being thinned to keep the frame smooth.

Follow the streets

Links follow the road geometry a file provides and are aligned to the street network where it does not. Should neither be available, straight lines and a reference grid make the fallback plain rather than suggesting a precision the data cannot support.

Inspect any element

Choose a node or a link and its whole record opens in one place: position, type, cluster, connections, lengths, received power and every other value the dataset supplies, each shown as the file states it, with nothing inferred in its place.

Keep data local

A file is read within the browser or the desktop application and is never sent to Optravo. Only the map area is requested from the map and routing services, so the network stays with you until you decide to share it, with no upload to wait for.

Visualization performance across tested systems.

Workload
600,000 nodes and 1,203,000 links.
Runs
Each system completed ten 60-second runs of 3,600 frames.
Results
Median, p95 and p99 of the pooled 36,000 frames. Lower CPU frame cost is better.
Build
Built from production React 19.3.0, with strict mode and power saving disabled.

HP OMEN 40L

Ryzen 9 5900X / RTX 3060 12 GB

Median10.0 ms
p9516.8 ms
p9924.7 ms
CPU
AMD Ryzen 9 5900X (12C/24T)
GPU
NVIDIA RTX 3060 (12 GB GDDR6)
Memory
32 GB DDR4-3200
OS
Windows 11 Pro 26H2 Build 26300.9550
Driver
617.14 WHQL
Browser
Chrome 154.0.8037.92
Display
3840 × 2160
Build
React 19.3.0, production
Test conditions
Strict mode off; power saving off
Test date
2026-11-01
Workload
600,000 nodes, 1,203,000 links
Run-to-run variance
median ±8% across 10 runs; p95−median spread 6.8 ms

Apple MacBook Pro 14″ (M4 Pro)

M4 Pro 12-core / 16-core GPU

Median8.8 ms
p9514.2 ms
p9922.3 ms
CPU
Apple M4 Pro (12-core)
GPU
Apple M4 Pro 16-core GPU (unified memory)
Memory
24 GB unified
OS
macOS Golden Gate 27.0
Driver
Apple Metal 3
Browser
Safari 27.0
Display
3024 × 1964
Build
React 19.3.0, production
Test conditions
Strict mode off; power saving off
Test date
2026-11-03
Workload
600,000 nodes, 1,203,000 links
Run-to-run variance
median ±7% across 10 runs; p95−median spread 5.4 ms

Lenovo ThinkPad E15 Gen 4

Core i7-1255U / Iris Xe

Median12.8 ms
p9522.2 ms
p9930.9 ms
CPU
Intel Core i7-1255U (10C/12T)
GPU
Intel Iris Xe 96 EU (shared memory)
Memory
16 GB DDR4-3200 (8 GB soldered + 8 GB SO-DIMM)
OS
Windows 11 Pro 26H2 Build 26300.9550
Driver
Intel 32.0.101.7092
Browser
Firefox 157.0
Display
1920 × 1080
Build
React 19.3.0, production
Test conditions
Strict mode off; power saving off
Test date
2026-11-05
Workload
600,000 nodes, 1,203,000 links
Run-to-run variance
median ±9% across 10 runs; p95−median spread 9.4 ms

Lenovo Legion Tower 7i Gen 8

Core i9-13900K / RTX 4090 24 GB

Median7.6 ms
p9511.8 ms
p9920.3 ms
CPU
Intel Core i9-13900K (24C/32T)
GPU
NVIDIA RTX 4090 (24 GB GDDR6X)
Memory
64 GB DDR5-6000
OS
Windows 11 Pro 26H2 Build 26300.9550
Driver
617.14 WHQL
Browser
Chrome 154.0.8037.92
Display
3840 × 2160
Build
React 19.3.0, production
Test conditions
Strict mode off; power saving off
Test date
2026-11-07
Workload
600,000 nodes, 1,203,000 links
Run-to-run variance
median ±6% across 10 runs; p95−median spread 4.2 ms

Dell Latitude 7490

Core i5-8250U / UHD 620

Median15.2 ms
p9528.6 ms
p9940.0 ms
CPU
Intel Core i5-8250U (4C/8T)
GPU
Intel UHD 620 (shared memory)
Memory
16 GB DDR4-2400
OS
Windows 11 Pro 26H2 Build 26300.9550
Driver
Intel 31.0.101.2145
Browser
Edge 154.0.4258.37
Display
1920 × 1080
Build
React 19.3.0, production
Test conditions
Strict mode off; power saving off
Test date
2026-11-10
Workload
600,000 nodes, 1,203,000 links
Run-to-run variance
median ±11% across 10 runs; p95−median spread 13.4 ms

Alienware Aurora (ACT1250)

Intel Core Ultra 7 265KF / GeForce RTX 5070 12 GB

Median8.1 ms
p9512.7 ms
p9921.1 ms
CPU
Intel Core Ultra 7 265KF (20C/20T)
GPU
NVIDIA GeForce RTX 5070 (12 GB GDDR7)
Memory
32 GB DDR5-5200
OS
Windows 11 Pro 26H2 Build 26300.9550
Driver
617.14 WHQL
Browser
Chrome 154.0.8037.92
Display
3840 × 2160
Build
React 19.3.0, production
Test conditions
Strict mode off; power saving off
Test date
2026-11-13
Workload
600,000 nodes, 1,203,000 links
Run-to-run variance
median ±7% across 10 runs; p95−median spread 4.6 ms
Interaction and memory benchmark results for the 600,000-node workload on the HP OMEN 40L
Pan and zoomMain-thread interaction cost during sustained camera movement with the complete scene resident, measured as input-event dispatch to the requestAnimationFrame callback of the last synchronous frame, at a 60 Hz target.4.6 ms pan and 5.9 ms zoom median; p95 7.7 ms pan and 10.3 ms zoom
Filter applicationVisible layer filtering without scene graph reconstruction or layer reload.10.7 ms median; 18.4 ms p95
Identifier searchResolution and display of the first matching network element.12.2 ms median; 23.4 ms p95
Selection-to-frame-submitElement-selection event to the requestAnimationFrame callback that submits the resulting frame.6.3 ms median; 10.6 ms p95
View resetReturn to the overview camera state from a focused view.5.4 ms median; 9.0 ms p95
Extended camera zoom sequenceTwo 90-second scripted camera passes across a wide zoom range with the complete topology resident.93.9% within 16.67 ms; 100.5 ms maximum
App-tracked allocation (300 s)JavaScript heap and GPU-buffer use during a 300-second steady-state measurement after the complete 600,000-node scene had been loaded and warmed up.JavaScript heap remained within ±20 MB of the post-warm-up baseline; tracked GPU-buffer allocation remained within ±5 MB. No sustained upward trend was observed during the measurement. GPU-buffer bytes are tracked by the app allocator, not read from VRAM.

Blog

No posts yet

New research summaries are coming soon.

Research software for inspectable optical-access-network models.

Optravo is developed at the Chair of Communication Networks, University of the Bundeswehr Munich, as research software for inspecting large optical-access-network graph models. The software emphasizes provenance, stable identifiers, explicit topology and supplied metadata to support reproducible network analysis.

Sebastian GartnerComputer Science ExpertSubject Area: Software Development
Ritanshi AgarwalMaster of ScienceCommunications and Electronics Engineering

Request student access.

Optravo student access is intended for University of the Bundeswehr Munich (UniBw M) students using the platform for coursework, projects or a thesis.

Direct inquiries may also be sent to info@optravo.org.

Tell us about your studies at UniBw M

For University of the Bundeswehr Munich (UniBw M) students. Share your study level, degree program, and planned use of Optravo.

Required fields