Mobility Setup¶

The mobility scenario moves agents over the deployed topology and charges each cell change. It is not a separate topology: it runs on top of the two-tier architecture, because the depth-two and path-stretch results both need a serving edge UPF distinct from the session anchor.

graph LR
    UE([UE]) -- NR interface (closest) --> gNB[gNB]
    gNB -- N3 interface (closest) --> EUPF[Edge UPF]
    EUPF -- N9 interface --> PSA[PSA / anchor]
    UE -. moves .-> UE2([UE at t+1])
    UE2 -- NR interface (new closest) --> gNB2[gNB']

At every update interval each agent's position advances, the nearest gNB is recomputed, and a cell change is classified by whether the serving edge UPF changed. That classification is the only input to the charge; see Mobility Sigma Accounting.

Configuration¶

Mobility is disabled by default, which leaves the stationary always-on eMBB behaviour of the other scenarios untouched. Enable it in config.toml:

[simulation]
scenario_mode = "two_tier"   # mobility results need an anchor distinct from the edge
num_centralized_upfs = 5     # PSAs

[simulation.mobility]
enabled = true
model = "random_waypoint"   # Options: "none", "random_waypoint"
update_interval = 1.0       # Simulation seconds between position updates and HO checks
speed_kmh = 5.0             # Pedestrian ~5, vehicular ~50, HSR ~250
pause_time = 0.0            # Reserved for future stateful waypoint model
max_jump_km = 1.0           # Cap per-tick displacement (sanity bound)

update_interval is the handover-detection granularity, not a cosmetic setting. A move that crosses two cells inside one interval is seen as one handover, so a coarse interval under-counts events at high speed. The sweeps use 1.0 s.

Mobility Models¶

Random waypoint picks a destination and walks toward it, which is adequate while the displacement per tick stays small relative to the cell size. Gauss-Markov keeps a correlated heading, with alpha the memory coefficient, so a 120 km/h agent travels like a vehicle on a road instead of reversing direction at random. Update equations are in Mobility Models.

Running the Sweeps¶

The published results come from runner scripts that build the mobility config in code, each dispatched through main.jl:

julia --project main.jl national_sweep                # 27 targets x 3 profiles = 81 runs
julia --project main.jl national_sweep france,canada  # country subset
julia --project main.jl national_sweep all 2000 300   # smoke: 2000 agents, 300 s

The three profiles are fixed in runs/national_sweep.jl:

const MODELS = [
    ("pedestrian", 5.0,   () -> RandomWaypoint(5.0, 0.0, 2.0)),
    ("urban",      50.0,  () -> RandomWaypoint(50.0, 0.0, 20.0)),
    ("highway",    120.0, () -> GaussMarkov(120.0, 0.85, 5.0)),
]

A fifth argument filters to one profile, so the 81 cells shard into independent jobs and wall time tracks the longest single job rather than the sum of all of them:

julia --project main.jl national_sweep spain 0 1200 results/spain-highway.csv highway

Related runners: anchor_sweep varies the PSA count, which is the unpublished parameter the path-stretch result depends on, and federation and ntn cover the multi-operator and non-terrestrial boundaries.

Targets¶

A target is a (country, field, operator) triple. Field is the base-station source: the crowdsourced OpenCelliD field for every country, plus the official national registry where one exists. Operator ids are real MNCs, so the same operator can be compared across both fields, which is what makes the source dependence measurable rather than assumed.

The FCC ASR field is a single all-structures target rather than a per-operator split, because the registry records structures rather than licensee-tagged cells. France and Canada carry every operator in both fields, which is what makes their two-field comparison a controlled one.

Deployment parameters follow runs/national.jl: edge UPFs are second-level administrative units above 50 000 inhabitants, PSAs are round(population / 10M) clamped to \([2, 5]\), and the agent count is population times adoption divided by the scale factor. Because the edge UPF count comes from the administrative partition rather than from a tuned parameter, \(\beta\) is inherited from how each country subdivides itself, which is why it varies by a factor of thirty across the six countries while the mobility profiles stay identical.

Output¶

Each run appends one row per (target, profile) to results/national-sweep.csv, carrying per-depth event counts, byte totals for both architectures, core writes, and anchor distances. Column meanings and how to read them are on the analysis page.