Trying to assess the difference in performance across languages/platforms of a fairly standard and simple economic ABM.
A simple econophysics-like model: random agent pairs exchange random sums of money. The structure of the models is extensible: any number of assets can be added.
I'm implementing two versions of the same model: 'centralized'; 'decentralized'.
In this version there is a single class (central register) that holds all the information about the agents. Agents themselves call the 'central register' to know and manipulate their holdings.
In this version each agent keeps account of his holdings. A 'central register' class exists for the sole purpose of holding a list of references of all existing agents and items. Agents may call the 'central register' to get those lists.
The model is implemented in the same way (as best as I can) across different languages/platforms.
Num agents: 10000; Time steps: 1000:
Pyhton: 2082" (or 34.7') (347x)
Java: 59" (9.83x)
C++: 6" (1x)
Num agents: 100000; Time steps: 1000:
Python: NA
Java: 20"
C++: NA
The difference between languages is significant and can impact the feasibility of some scientific endeavors (like calibration). A decentralized version seems to be much faster than a centralized one (due to the sparsity of the matrix that keeps the accounts for all agents-items).