This is my 4th year thesis on "Agent based models of developer interaction in large scale software ecosystems". The model calculates the variation of some network characteristics like Connection, Separation with respect to time. The software development industry is now in full bloom. With the influx of multi-million dollar enterprise-level projects, like ChatGPT, or Autonumous Driving Cars, huge networks of developers are becoming more and more common behind such projects to drive them to completion. We aim to understand these complex social processes for better disemmination of information between developers, identify bottlenecks in flow of information, allocation of resources and ease of governance.
A promising apporach to understand such networks is to see if we can simulate them properly. Creating such simulations can be done using the agent based modelling paradigm. One tool to the implement said Agent Based Modelling paradigm is netlogo.
We assume that a software development ecosystem is divided into many modules, on which the development team works. Each team will also have a team lead who is in charge of co-ordination between the modules as they talk to the other team leads. The developers only interact amongst their team mates. Between the team leads we assume that communication happens following only these three structures as depicted above, a ring, a fully connected graph and a star network where the team leads report to a CEO. Ultimately we have proposed three models following these three structures of communication.
Netlogo. This is the agent-based modelling software used to simulate our models. The models have five parameters, each that control how the networks grow and evolve over time:
- Rate-of-Connection
- Number of people in the team
- Number of teams that exist
- Number of people to be added to be th team after each tick.
- Number of random connections between teams The Netlogo interface for one of the models is present down below.
Datasets used: Openstack, Eclipse, Android
The outputs from these models have then been scaled by multiplying them with a scalar factor as real-world data is apt to be more dense than the what we have simulated. Using these parameters, we have built models that show high congruence with characteristics that are shown by real world data obtained from software development projects like Openstack, Eclipse, Android. The extensive validation on three different datasets of three different types of software development projects ( a Cloud platform, an IDE, and a mobile OS) lends more solidarity towards these models being closer to reality.
To work with the idea of social networks in computer science we abstract these networks out as graphs with nodes being the the developers and interactions between them are presented as edges. The variations of the network characteristics Connection and Separation with time vs with what is observed in the real world is given below. The first graph is that of the metric Connection, which is ratio of ( 2* the number of edges) and (the number of vertices) in the network. Our simulated graph, in accordance to the empirical graph also 3 regions which include a gentle rise followed by a steady plateau, which is then followed by the third stage with an exponential rise. The graph to the right depicts the curve of Separation which has the two distinct regions - a sharp rise and fall. The resemblance in both the cases is striking!
Netlogo spills out csv data while running experiments on model parameters. However, said data is very messy. This is a UI, built on streamlit and deployed on the streamlit cloud at Behavioural Space Parser that can appropriately parse the csv files. The code is available in the Helper Tool folder. It can be downloaded and running the following at your terminal will get it fired up.
pip install -r requirements.txt
streamlit run application.py
