Inconsistent Volume Counts Across Simulations with Fixed Random Seed

[OliBus801]

Hi everyone,

I'm encountering some inconsistencies in MATSim simulation results, even when using a fixed random seed. I've been trying to reproduce a specific simulation run in terms of volume count, but I'm getting varying results.

Detailed Observations:

• Initially, I observed significant variability in results despite using a fixed random seed.
• To investigate, I examined the configuration and disabled multi-threading across QSim, general, and eventManager settings. This change led to a noticeable reduction in result variability.

Example:
I ran 10 simulations of 50 iterations each on the SiouxFalls-2014 scenario while maintaining an identical configuration to a baseline run. Importantly, I ensured the random seed was reset before each simulation. The results showed:

• 6 simulations matched the volume counts of the baseline run exactly.
• 4 simulations exhibited a Mean Relative Error of approximately 12%, with these 4 simulations having identical results among themselves.

Question:

I'm curious about the potential causes of these discrepancies. Could CPU architecture be a factor? Additionally, are there any recommended techniques to guarantee identical volume counts across simulations?

Attached information:

Output Configuration file used across all simulations:
output_config.txt

Discrepancy Graph (X-axis: Simulation Number):