Skip to content

Layout Investigation — device/label overlap in generated labs

Status: investigation only (tools/, src/ untouched). Scope: JASDF_AEWC_JADGE.pkt (290 devices) and JASDF_AEWC_JADGE_standard.pkt (199 devices) from filedrop/uploads/, the generator source in JASDF_AEWC_source.zip, tools/build_large_topology.py + tools/lab_api.py, and 15 human-authored corpus files from ../corpus/internet/ (index: ../corpus/CORPUS_INDEX.tsv, decode_status == OK).

All coordinates below are WORKSPACE/LOGICAL/{X,Y} workspace units. Calibration note: at Packet Tracer's default (100 %) zoom, 1 workspace unit ≈ 1 screen pixel — verified against JASDF_AEWC_packet_tracer_verified.png, where the airframe pitch of 310 units measures exactly 310 px between icon centres.

1. Measurements of the generated labs

1.1 Global numbers

Metric JASDF_AEWC_JADGE.pkt (enterprise) ..._standard.pkt
devices with logical coords 290 (0 missing) 199 (0 missing)
bounding box X [60, 3700], Y [130, 5140] X [60, 3700], Y [130, 5140]
span 3640 × 5010 units 3640 × 5010 units
density 1 device / 62 914 units² 1 device / 91 684 units²
busiest 500×500 window 23 devices at (510, 130) — airborne block 17 devices at (250, 1550) — CRC/server block
name length min 7 / median 14 / mean 13.4 / max 23 min 7 / median 13 / mean 13.3 / max 23
exactly coincident pairs (d=0) 0 0

1.2 Pair counts at centre distance (icon is ~40–60 units wide)

threshold JADGE (290 dev) standard (199 dev)
≤ 30 0 0
≤ 40 (icon size) 12 3
≤ 50 14 5
≤ 60 15 6
≤ 80 21 9
≤ 100 (below human minimum) 56 18
≤ 120 (human "comfortable" floor) 84 31
≤ 150 171 87

Nearest-neighbour distance per device (JADGE): min 30.4, p10 58.3, median 135.0, mean 139.6, max 403.1. 24 of 290 devices have a neighbour within 40 units (standard: min 30.4, p10 85.0, median 150.3 — 6 of 199 within 40).

Axis-pitch histograms show the hard-coded constants: the most common X gaps are 10 (×34), 30 (×12), 5 (×8), 20 (×8); the most common Y gaps are 180 (×6), 190 (×4), 70 (×3).

1.3 Worst clusters (connected components at d ≤ 40)

JADGE: 12 clusters, every one a pair (top 10 shown; all 12 are listed since they are all equal size):

# devices coordinates d
1 TFTP-SVR / CRC-CENTRAL-RTR-01 (ISR4331) (655, 1910) / (650, 1880) 30.4
2 TDL-GW-03 (2911) / NOSC-TAP-SNIFFER (Sniffer) (2310, 1150) / (2300, 1120) 31.6
3 MISSION-WEB-SVR / CRC-CENTRAL-RTR-02 (ISR4331) (830, 1910) / (840, 1880) 31.6
4 E767-02-WLAN-AP / E767-01-WC-PC (510, 305) / (550, 305) 40.0
5 E767-03-WLAN-AP / E767-02-WC-PC (820, 305) / (860, 305) 40.0
6 E767-04-WLAN-AP / E767-03-WC-PC (1130, 305) / (1170, 305) 40.0
7 E2D-01-WLAN-AP / E767-04-WC-PC (1440, 305) / (1480, 305) 40.0
8 E2D-02-WLAN-AP / E2D-01-WC-PC (1750, 305) / (1790, 305) 40.0
9 E2D-04-WLAN-AP / E2D-03-WC-PC (510, 575) / (550, 575) 40.0
10 E2D-05-WLAN-AP / E2D-04-WC-PC (820, 575) / (860, 575) 40.0
(11) E2D-06-WLAN-AP / E2D-05-WC-PC (1130, 575) / (1170, 575) 40.0
(12) E2D-06-WC-PC / E2C-01-WLAN-AP (1480, 575) / (1440, 575) 40.0

standard.pkt keeps only clusters 1–3 (the enterprise profile adds the airborne block and the NOSC/TDL rows).

Near-overlaps at d ≤ 80 (would still visibly touch): TDL-GW-02/NOSC-CYBER-OBS (42.4), AAA-SVR/CRC-NORTH-RTR-02 (50.0), TDL-GW-01/NOSC-NMS-SVR (58.3), SYSLOG-SVR/CRC-NORTH-RTR-01 (62.6), SATCOM-GW-01/NOSC-SW-01 (70.7), PKI-CA-SVR/CRC-NORTH-DSW (72.1).

1.4 Icon and label collisions

Icon artwork sizes measured from JASDF_AEWC_packet_tracer_verified.png (1:1 with workspace units): router cylinder ≈ 50×40, L2/L3 switch ≈ 46–50, AccessPoint-PT ≈ 48×34, PC-PT ≈ 34×40, Server-PT tower ≈ 28×48. PT draws the model caption above the icon and the device name below it, i.e. two text lines per device. Label font measures ≈ 6.3 px/char (E767-02-WLAN-AP = 15 chars ≈ 95 units wide).

AABB icon overlap (icons treated as squares centred on X,Y), JADGE:

icon box overlapping pairs
40 units 4
50 units 16
60 units 20
70 units 21

Label collisions (label box = 6..9 px/char × len(name), drawn below the icon):

char width label↔label pairs (JADGE) labels over another icon (JADGE) label↔label (standard) over icon (standard)
5 px 9 29 0 10
6–7 px (measured ≈ 6.3) 9–12 29–30 0–1 10
8 px 23 31 9 11
9 px 29 38 13 15

The PT screenshot confirms this visually: in the airborne row E767-02-WLAN-AP and E767-01-WC-PC are printed on top of each other (centres 40 units apart), as are E767-03-WLAN-AP/E767-02-WC-PC and the next pair. Long names make it worse: CRC-CENTRAL-DISPLAY-SVR (23 chars ≈ 155 units wide) sits only 190 units from CRC-CENTRAL-TRACK-SVR (21 chars), so the two labels nearly touch even where icons are clear. The seed/lab default name Power Distribution Device0 (27 chars ≈ 185 units) is the worst case in the wider corpus.

Bottom line: 12 device pairs physically overlap at icon scale, 24 devices have a neighbour inside icon distance, ~29 device labels collide with icons and 9–29 label pairs collide with each other (screenshot-verified), and 56/84 pairs fall below the human-authored minimum/comfortable spacing (100/120).

2. Root cause

One sentence: there is no layout algorithm at all — every generator writes hand-typed literals or index-derived constant-pitch formulas straight into lab_api.Lab.add_device(model, name, x, y), which blindly stores whatever coordinates it is given, so independent section grids overlap wherever their hard-coded origins coincide and nothing ever checks icon size, label length, device kind, occupancy, or canvas bounds.

Details:

  1. tools/lab_api.py is layout-blind. add_device(..., x: int = 100, y: int = 100, ...) (lab_api.py:259) patches the template XML's <WORKSPACE><LOGICAL><X>/<Y> with the caller's numbers (lab_api.py:277-278) and does nothing else. Defaults mean that omitting coordinates stacks every device at (100, 100). The only position bookkeeping is self._pos += 6 (lab_api.py:270) — that is the physical rack slot X for PHYSICALWORKSPACE, unrelated to the logical canvas. There is no occupancy map, no collision test, no clamping.

  2. Coordinates are hand-assigned / index-derived, spacing is constant per section. In JASDF_AEWC_source.zip → build_jasdf_aewc.py:

  3. airborne block: bx, by = 200 + ((ac-1)%6)*310, 130 + ((ac-1)//6)*270 (line 438); within an airframe MR at (bx,by), SW/MC at bx+175, AP at (bx, by+175), WC-PC at bx+350 (line 456). The intra-airframe offset 350 exceeds the inter-column pitch 310, so airframe N's WC-PC lands 40 units from airframe N+1's WLAN-AP — clusters 4–12 above.
  4. server farm: x = 130 + (i%5)*175, y = 1650 + (i//5)*130 (line 289-290).
  5. CRC block: x = 250 + i*400 (line 348), routers at y=1880, RTR-02 at x+190, DSW at x+95, 2100, servers y=2320, consoles x-60+j*130, 2500.
  6. NOSC row: servers at 1850/2000/2150/2300, y=1120 (lines 330-339); TDL gateways 2050+(i-1)*130, y=1150 (line 241). The two grids are only 30 units apart in y and overlap in x → clusters 1–2.
  7. radar sites bx = 120 + (idx%8)*310 (line 380), airbases bx = 120 + (idx%5)*430 (line 404) — the same constant-pitch pattern.

  8. tools/build_large_topology.py is identical in spirit: x = 100 + (site-1)*420 + (branch-1)*190 + (n%2)*70, y = 760 + (n//2)*130 (lines 263-264), IoT 1450+(i%2)*120, 80+(i//2)*130 (line 284). Client columns are 70 units apart — icons clear, labels (LT-S11-1, 8 chars ≈ 60 units) almost touch.

  9. Nothing accounts for icon size, label length, or device kind. All spacings (70, 95, 120, 130, 150, 175, 190, 200, 250, 270, 310, 400, 430…) are magic numbers. A 2911 router (~50 units wide + 2 label lines) and a PC-PT (~34 units) get the same pitch; a 23-char name gets the same cell as RDR-1.

  10. Section grids share one canvas without a global occupancy model. build_jadge_dc's server grid (y 1650–2170, x 130–830) overlaps build_crcs' routers (y 1880, x 250/650/…); build_hq's NOSC row overlaps build_backbone's TDL row. Each build_*() method picked its own "empty" region by eye; nothing reconciles them.

  11. Workspace range: nothing keeps devices inside anything. The file format stores no canvas extent (no viewport/zoom/extent element in the XML); PT's logical workspace pans and zooms freely. Empirically PT's stock page is about 3900 × 3900 units — the default Power Distribution Device0 sits at (3898, 3900) in tools/network_seed.xml and in corpus labs, and human labs keep content within ≈ (84…2538) × (29…1115) in the sample (well inside that page). The generated JADGE lab runs to Y = 5140, i.e. past the default page: users must scroll to find the airbase rows, and no generator notices.

3. Geometry of real (human-authored) corpus labs

Method: stratified sample of 15 files with decode_status == OK from ../corpus/CORPUS_INDEX.tsv (865 OK of 1059), decoded with pkt_core; PT's auto-placed Power Distribution Device* excluded (it is an artefact parked at the (3898, 3900) corner in most files and would otherwise fake a 0-distance neighbour). 4 files in the first sampling pass could not be XML-parsed (not well-formed (invalid token) inside .pka activity XML) and were replaced.

file n X range Y range NN min NN p25 NN median pairs ≤ 40
imsiddhant …EtherChannel 3 464–896 162–258 188.0 188.0 188.0 0
hackthacker …RING TOPOLOGY 8 508–1188 189–713 85.1 89.3 108.7 0
Jawadyyy …IotServer 6 84–528 44–319 117.2 117.2 136.0 0
drespencer …Static Routing (P2) 16 88–897 69–372 70.6 83.7 97.1 0
blup-blup-spec …VLANbsic 21 94–908 69–421 69.0 72.3 84.5 0
MIETDevelopers …startopology 6 484–1110 102–564 191.0 191.0 291.0 0
smsufi …ACL-extended 12 164–983 154–516 131.6 134.6 157.9 0
hackthacker …BUS topology 8 450–1056 278–478 187.0 187.2 188.6 0
Rakibul73 …10.pka 9 106–692 50–336 105.5 108.6 111.4 0
drespencer …DNS 9 640–1014 58–348 93.0 100.7 112.4 0
imsiddhant …Vlan Lab 8 484–1230 225–608 130.0 132.0 144.3 0
vaamonde …Cenario-A 10 348–1056 62–389 125.7 141.9 147.3 0
dishavp …CHAP 18 107–2538 113–1115 50.2 50.2 50.2 0
dishavp …p3sub,ospf 25 295–2008 29–922 81.3 88.4 146.4 0
kristof-levai …HSRP_exercise1 8 224–534 69–362 61.8 72.0 111.1 0

Aggregate over 167 devices / 15 files:

  • nearest-neighbour distance: min 50.2, p5 50.2, p10 70.6, p25 88.4, median 125.7, p75 172.1, p90 203.5, max 307.1, mean 134.3
  • pairs within 40 units: 0 (0 %) — humans never place icon centres at icon distance; the tightest real case is 50.2 units
  • NN ≤ 60: 7.2 %, ≤ 80: 13.8 %, ≤ 100: 31.7 %, ≤ 120: 47.3 %, ≤ 150: 67.7 %, ≤ 200: 89.8 %

Interpretation for the fix: comfortable spacing ≈ 90–180 units between centres; the hard floor is ~80 (label/icon clearance), and anything ≤ 60 reads as "overlapping" to a human. The generated labs violate even the 50-unit observed minimum. Human layouts also stay near the top-left of the ~3900×3900 PT page and use whole-coordinate placement with pitch 80–200 — exactly the range the generator's constants (70/95/130/175) fall under.

A second independent stratified pass over 15 further files (81 devices, PDD excluded) confirmed the same picture: min 50.2, p25 85.8, median 146.4, mean 150.7, and again 0 pairs within 40 units.

4. Proposed fix

4.1 Sizing model (the missing primitive)

Give every device a bounding box that includes its labels, per kind — measured icon sizes (§1.4) plus the two label lines PT draws (model caption above, name below):

CHAR_W  = 7      # units per label char (6.3 measured; 7 leaves safety margin)
GUTTER  = 28     # clearance added around every box
ICON_W  = {  # width x height of icon artwork, default zoom
    "router": (52, 44), "switch": (50, 42), "multilayer-switch": (50, 42),
    "access-point": (50, 36), "pc": (36, 42), "laptop": (36, 42),
    "server": (30, 48), "iot": (40, 40), "sniffer": (50, 40),
    "firewall": (52, 42), "pdd": (40, 40), "default": (52, 44),
}

def device_box(kind, name, model_caption):
    w = max(ICON_W[kind][0], CHAR_W * len(name) + 16, CHAR_W * len(model_caption) + 16)
    h = ICON_W[kind][1] + 2 * 18          # model line above + name line below
    return w + GUTTER, h + GUTTER

Worked examples: PC-PT / E767-01-WC-PC (12 chars) → 100×108; ISR4331 / CRC-CENTRAL-RTR-01 (18) → 154×100; Server-PT / CRC-CENTRAL-DISPLAY-SVR (23) → 189×112; Power Distribution Device0 (27) → 217×96. This is why a uniform grid alone cannot work: a router with a long name needs ~2× the cell of a PC.

4.2 Placement strategy: greedy row packing with occupancy test

Deterministic and cheap (290 devices ≈ trivial); no force-directed randomness (bad for reproducible byte-exact .pkt output, which this repo's round-trip tests require):

  1. Group devices semantically (site / tier / airframe — the generators already know these). Lay groups out as horizontal bands, top to bottom, left-aligned; keep the group's intended rough position via an anchor (anchor("airborne", x=200, y=130)) so topology geography survives.
  2. Inside a band, place devices row-major: advance a cursor by the device's device_box width; before committing, AABB-test against every already-placed box and skip forward until free (greedy packing that skips occupied cells). On band-width overflow, wrap to the next row of the band with row height = max box height in that row.
  3. Minimum pitch rule (validated against the corpus): min_centre_distance = max(70, w_a/2 + w_b/2) where w are the label-inclusive box widths — i.e. defaults: column pitch 200 for names ≤ 16 chars, 260 for 17–24, 320 for ≥ 25 chars; row pitch 150. Never place two devices closer than 80 units centre-to-centre under any circumstance (corpus floor is 50, comfortable median 126).
  4. Canvas containment. Target region X ∈ [60, 3860], Y ∈ [60, 3860] (PT's ~3900×3900 default page, per the seed's Power Distribution Device0 at (3898, 3900)). Before placing, compute total band height; if it overflows, wrap to a second page column at x_offset += 3900 (keeping the minimum pitch) rather than shrinking spacing. A final clamp() pass moves any device whose label box would exit the region back inside. At 290 devices and 200×150 cells the whole JADGE enterprise lab needs ≈ 2200×2200 units — it fits on one page comfortably (today it needs 3640×5010 because clusters are clumped and the rest is empty).
  5. Optional refinement for cabling readability: after packing, a light relaxation (move a device at most 1 cell toward the centroid of its link neighbours, re-checking AABBs) — but this is a nicety, not required to fix overlap.

4.3 Where it plugs in

  • tools/lab_api.py — the right home, because every generator (build_large_topology.py, build_wired_lab.py, build_configured_lab.py, make_scored_lab.py, and the JASDF kit) already funnels through Lab.add_device. Add a Layout helper (sizing model + occupancy grid + clamp) and change the contract to add_device(..., x=None, y=None, group=None) where None means "ask the planner". Explicit x,y keeps working for tests and for the seed.
  • The generators then stop passing magic numbers: they pass their semantic group (group="airframe:E767-01", "crc:CENTRAL", "servers") and optionally an anchor; the formulas at build_jasdf_aewc.py:241,289,348,438 and build_large_topology.py:263-264 collapse into group declarations. (build_jasdf_aewc.py lives in JASDF_AEWC_source.zip, not in this repo — it would need the same change in its JASDF.add() wrapper, which is a one-liner since it already delegates to lab.add_device.)
  • A checker, e.g. tools/check_layout.py (or a pkt_core test), recomputing §1's metrics (pairs ≤ 40, pairs < 80, label AABB collisions, bbox ≤ 3900) from a decoded file so CI can fail on regressions.
  • Algorithm: greedy row-major packing with per-device label-inclusive boxes and an occupancy/skip-occupied-cells test, grouped into anchored bands, wrapped into 3900×3900 pages.
  • Spacing: cell width max(icon_w, 7·len(name)+16, 7·len(model)+16) + 28; column pitch 200 (typical), 260 (long names), 320 (≥25 chars); row pitch 150; hard floor 80 units centre-to-centre — matching human corpus practice (min 50, p25 88, median 126) and clearing the largest icon (~50) plus label gutters.

5. Method notes / reproducibility

  • Decoding: pkt_core.decode() from src/; devices are NETWORK/DEVICES/DEVICE, name ENGINE/NAME, kind ENGINE/TYPE, coordinates WORKSPACE/LOGICAL/{X,Y}.
  • Measurement scripts were kept outside the repo (/tmp/layout_measure.py, /tmp/layout_detail.py, /tmp/corpus_final.py); nothing under tools/ or src/ was modified.
  • Screenshot calibration source: JASDF_AEWC_packet_tracer_verified.png (940×470) showing the airborne row at 1:1 zoom; the visible E767-02-WLAN-AP / E767-01-WC-PC label pile-up matches cluster #4.
  • Corpus failures (XML not well-formed inside some .pka activities) were excluded and replaced in the sample; they are a decode edge case unrelated to layout.