Distributions and mobility

Why this scenario exists

A constant workload flatters a placement policy. Every task of a real edge deployment arrives at its own instant, carries a different payload, and asks for a different amount of CPU, and a moving device changes the access link it pays for. Papers that sweep only constants therefore measure the policy under one point of the workload space.

This scenario runs the same topology twice, first with the file values and then with those values replaced by their stochastic counterparts, so the difference between the two rows is the effect of randomness alone. It is also the demonstration that the configuration can be built from Python: the second run never edits a file, it rewrites the objects returned by load_settings.

What it shows

How to configure a scenario from Python instead of editing the settings files. The settings folder describes constants: every task of an application has the same length, the same sizes, and the same gap between requests, and the devices stand still. This example loads those files, then replaces the constants by probability distributions and lets the devices move.

The distributions live in the pureedgesim package and are also accepted inside the XML files. The example uses them directly:

flowchart LR
    length["task_length<br/>constant MI"] --> lognormal["LogNormal<br/>mean, std = mean / 3"]
    size["request_size<br/>constant KB"] --> weibull["Weibull<br/>shape 1.5, scale = mean / 0.9"]
    rate["rate<br/>requests per minute"] --> exponential["Exponential gap<br/>60 / rate seconds"]
    still["mobility false"] --> gauss["gauss-markov<br/>alpha 0.8, std 15, speed_std 0.4"]
    gauss --> speed["speed<br/>Normal(2.0, 0.5) bounded 0 to 5"]

• LogNormalDistribution for the task length, with the same mean as the constant and a third of it as standard deviation;
• WeibullDistribution for the request size;
• ExponentialDistribution for the inter-arrival gaps, derived from the rate;
• NormalDistribution(2.0, 0.5).bounded(0.0, 5.0) for the speed, so a draw is never negative;
• the gauss-markov mobility model for every device type, with its alpha, std, and speed_std options.

The scenario

Both runs use examples/distributions_and_mobility/settings with 20 devices, five simulated minutes, the ALL architecture, and random_seed=7. The first run keeps the file values; the second replaces them as described above. Randomness.set_seed and Randomness.begin_scenario make the draws reproducible across the two runs.

The script also prints a few draws before running, so the shape of each distribution is visible without reading the CSV: one gap from the heavy application, and six lognormal task lengths.

Run it

From the repository root, after pip install -e .:

python examples/distributions_and_mobility/distributions_and_mobility.py

What to look for

HEAVY_COMP_APP: every task draws 30000 MI, 2500 KB, gaps of 11.1 s
draws from the lognormal task length: 38135, 18575, 18459, 33461, 38118, 20595
constants, static      tasks= 1035 ok=  751 delay=0.2025   failed(delay)= 284 failed(mobility)=   0 energy=11.5627
distributions, mobile  tasks= 1006 ok=  792 delay=0.2132   failed(delay)= 214 failed(mobility)=   0 energy=10.6488

• The task count changes (1035 to 1006): the exponential gaps produce a different number of requests in the same simulated time.
• Fewer tasks miss their deadline in the second run (284 to 214), so 792 succeed against 751, even though the mean delay is higher (0.2025 s to 0.2132 s), because the lognormal mix includes many short tasks: the average is dragged up by the long draws while the short ones clear their budget. The per-tier columns show where the gain is, the mist tier returns 507 successes out of 510 placements in the distributed run against 361 out of 448 in the constant one.
• failed(mobility) stays at zero in both rows: moving devices leave and re-enter range, but they are not lost, so mobility changes the destination, not the outcome.
• Results go to examples/distributions_and_mobility/output/, where both runs are appended to one Sequential_simulation.csv.

Replace the gauss-markov model by manhattan, levy-walk, or the other models listed in the README to compare movement patterns under the same workload.

Files

• distributions_and_mobility.py with the parameter overrides and both runs
• settings/ with simulation_parameters.properties, applications.xml, edge_devices.xml, edge_datacenters.xml and cloud.xml

Where to go next

• Custom orchestrator for a different kind of Python-side configuration, the placement policy
• The settings folder section of the main README for the distribution syntax inside the XML files