Font Size: 

Computational Social Science is one of the newest and most promising research lines trying to understand human behaviour. Within this context, Agent-Based Modelling is a tool increasingly being used, as it allows 1) to explicit human behaviour at an individual scale, and 2) to detect emergent properties of the system as a whole. An important characteristic of ABM is that social scientists seem more comfortable with this than with other tools (like classical equation based solutions), as ABM provides a friendly framework to integrate social science knowledge.
Despite the potential advantages of ABM, there is an ongoing discussion about the way these models should be used. The standard approach for ABM is the design of abstract, simple models with a low number of parameters and agents. Nevertheless, the access to increasing numbers of datasets and computing facilities has opened the door to highly complex and realistic models.
This presentation explores the problems and challenges that archaeological simulation modellers need to face if they are interested on exploiting the power of High-Performance Computing (HPC). Complex social scenarios require simulations with advanced artificial intelligence and large numbers of agents. HPCs seem the solution to tackle this added complexity, but little work has been done in order to understand real consequences of using ABM in HPC environments. These systems split a problem in different modules, each of them being executed on a computer node that should communicate with the other ones as little as possible. This idea is exactly opposite to ABM, where agents affect each other, share information, and modify the environment, so the ratio of communication between them is high.
Moreover, the analysis of simulation results is difficult, because scientists should be able to explore the model, both in terms of individual traits (adaptation, co-evolution) and emergent global processes (collectives, spatial patterns)
We present Pandora as a possible solution to these problems. It is a library specially designed to deal with distributed, fully scalable archaeology-related ABM. This C++/MPI library manages the distribution of executions in a transparent way in any system. Parallelization is automatically organized from spatial coordinates, and each computer node only needs to communicate with its neighbours. This tool is complemented by Cassandra, an application designed to analyse simulation results combining statistics, GIS and visualization.
Unfortunately there is no universal answer to the challenge of distributing ABM, and this paper shows how the choice of solutions should basically depend on the relation between time and space scales. We will discuss the issues generated from the use of Pandora in different case studies, developed between archaeologists, anthropologists and computer scientists. These experiences have been useful to detect some of the problems that will be common in large-scale ABM like parallelization of advanced agents, GIS integration, stochasticity of the system and potential bottlenecks. Conclusions describes how these issues can be fixed or avoided, in order to advance towards a generalised use of HPC in archaeological simulations.

Agent-Based Modelling; High-Performance Computing