Custom settings and a parallel sweep: six scenarios in two passes

What it shows

Two things that the other examples keep separate: a scenario that comes from a private settings folder instead of the default one, and a sweep over that scenario run twice, once in the main process and once in worker processes.

The settings folder declares a grid of six scenarios, three architectures by two algorithms, and the script runs that grid twice with the same seed:

parameters = load_parameters(SETTINGS / "simulation_parameters.properties")
parameters.parallel_simulation = parallel
parameters.random_seed = 42
simulation = Simulation(SETTINGS, output, parameters, load_settings(SETTINGS))
rows = simulation.launch()

The claim of the example is that parallelism is a scheduling decision, not a modelling one: the two passes must produce the same six rows. The script prints the two sweeps produced 6 rows each, identical: True when they do, and the test compares the rows column by column.

flowchart LR
    grid["settings/<br/>3 architectures<br/>x 2 algorithms"] --> mode{"parallel_simulation"}
    mode -->|false| seq["main process<br/>6 scenarios one after another"]
    mode -->|true| pool["ProcessPoolExecutor<br/>min(8, 6) workers"]
    pool --> w1["worker 1<br/>MIST_AND_EDGE / ROUND_ROBIN"]
    pool --> w2["worker 2<br/>MIST_AND_EDGE / TRADE_OFF"]
    pool --> w6["worker 6<br/>CLOUD_ONLY / TRADE_OFF"]
    seq --> csv1["Sequential_simulation.csv"]
    w1 --> csv2["Parallel_simulation.csv"]
    w2 --> csv2
    w6 --> csv2
    csv1 --> cmp["6 rows, all 56 columns equal"]
    csv2 --> cmp
    classDef out fill:#cfe8ff,stroke:#333;
    class csv1,csv2 out;

Each worker re-derives its own random stream: run_scenario calls Randomness.set_seed(parameters.random_seed) and then begin_scenario(iteration), so scenario i always draws from the seed mixed with i, whichever process runs it. That is why the two passes can be compared at all.

The scenario

settings/ holds one fleet of 100 devices, 10 simulated minutes, the six scenarios above, and random_seed= left empty, because an unseeded run is not reproducible. The script pins seed 42 for both passes so the comparison is possible; without a seed the two sweeps would differ for that reason alone, not because of the workers.

Run it

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

python examples/custom_settings_parallel/custom_settings_parallel.py

The same script runs without installing the package, as long as src is on the path:

PYTHONPATH=src python examples/custom_settings_parallel/custom_settings_parallel.py

What to look for

The console prints one line per scenario for each pass, then the wall time and the verdict:

settings folder: D:\Charef\doctorat\Work folder\PyEdgeSim\PureEdgeSim\examples\custom_settings_parallel\settings

sequential  MIST_AND_EDGE  ROUND_ROBIN  devices= 100 tasks= 10150 ok= 1013 avg_delay=0.6084   energy=47.4811
sequential  MIST_AND_EDGE  TRADE_OFF    devices= 100 tasks= 10150 ok= 1014 avg_delay=0.465    energy=47.8248
sequential  ALL            ROUND_ROBIN  devices= 100 tasks= 10150 ok= 1015 avg_delay=0.5916   energy=47.6758
sequential  ALL            TRADE_OFF    devices= 100 tasks= 10150 ok= 1016 avg_delay=0.5287   energy=49.1103
sequential  CLOUD_ONLY     ROUND_ROBIN  devices= 100 tasks= 10150 ok= 1015 avg_delay=0.1575   energy=58.8768
sequential  CLOUD_ONLY     TRADE_OFF    devices= 100 tasks= 10150 ok= 1015 avg_delay=0.1576   energy=58.5877
sequential  6 scenario(s) in 23.06 s wall time

parallel    MIST_AND_EDGE  ROUND_ROBIN  devices= 100 tasks= 10150 ok= 1013 avg_delay=0.6084   energy=47.4811
parallel    MIST_AND_EDGE  TRADE_OFF    devices= 100 tasks= 10150 ok= 1014 avg_delay=0.465    energy=47.8248
parallel    ALL            ROUND_ROBIN  devices= 100 tasks= 10150 ok= 1015 avg_delay=0.5916   energy=47.6758
parallel    ALL            TRADE_OFF    devices= 100 tasks= 10150 ok= 1016 avg_delay=0.5287   energy=49.1103
parallel    CLOUD_ONLY     ROUND_ROBIN  devices= 100 tasks= 10150 ok= 1015 avg_delay=0.1575   energy=58.8768
parallel    CLOUD_ONLY     TRADE_OFF    devices= 100 tasks= 10150 ok= 1015 avg_delay=0.1576   energy=58.5877
parallel    6 scenario(s) in 10.24 s wall time

the two sweeps produced 6 rows each, identical: True

Measured run

The six rows of output/<timestamp>/Parallel_simulation.csv, measured 2026-10-28 on seed 42, 100 devices, 10 simulated minutes each. The sequential CSV holds the same numbers.

Architecture Algorithm Tasks Successful Avg delay WAN time Request makespan Energy
MIST_AND_EDGE ROUND_ROBIN 10150 1013 0.6084 s 5.7086 s 4.4269 s 47.4811 Wh
MIST_AND_EDGE TRADE_OFF 10150 1014 0.465 s 6.7925 s 3.4962 s 47.8248 Wh
ALL ROUND_ROBIN 10150 1015 0.5916 s 15.3839 s 3.8451 s 47.6758 Wh
ALL TRADE_OFF 10150 1016 0.5287 s 30.9079 s 4.1262 s 49.1103 Wh
CLOUD_ONLY ROUND_ROBIN 10150 1015 0.1575 s 563.4337 s 2.87 s 58.8768 Wh
CLOUD_ONLY TRADE_OFF 10150 1015 0.1576 s 498.8261 s 2.87 s 58.5877 Wh

Reading the sweep:

Tune it

Files

Where to go next