Dans le domaine des systèmes multi-agents, la hausse constante du nombre d'entités implique un besoin en ressource de calcul de plus en plus important. Cependant, les nombreux outils et plates-formes permettant le développement de simulations multi-agents sur CPU ne supportent plus cette demande en perpétuelle augmentation. Une solution est de se tourner vers le calcul haute performance et notamment vers le calcul sur carte graphique : le GPGPU. Apportant un rapport performance/prix imbattable, cette technique souffre cependant d'une programmation très spécifique qui limite son adoption par la communauté. Dans cet article, nous faisons un état des lieux des simulations multi-agents sur GPU et identifions les solutions les plus prometteuses en vue d'une généralisation de l'utilisation de cette technologie dans la communauté multi-agent. ABSTRACT. In some application domains, using a Multi-Agent Systems (MAS) modeling approach may require to handle a large number of agents (crowds, traffic, ecosystems, etc.). In such cases, the computational resources which are needed often raise scalability problems. Considering this kind of issues, General-Purpose computing on Graphics Processing Units (GPGPU) appears to be an appealing solution as it enables huge speed up on a regular PC. However, this technology relies on a highly specialized architecture, implying a very specific programming approach. That is the reason why GPGPU is not widespread technology in the MAS community. This paper reviews the literature which is at the intersection between MAS and GPGPU. The different approaches used are presented and the most promising solutions for a generalization of GPGPU technology in our community will be highlighted.

Abstract. General-Purpose Computing on Graphics Units (GPGPU) is today recognized as a practical and efficient way of accelerating software procedures that require a lot of computing resources. However, using this technology in the context of Multi-Agent Based Simulation (MABS) appears to be difficult because GPGPU relies on a very specific programming approach for which MABS models are not naturally adapted. This paper discusses practical results from several works we have done on adapting and developing different MABS models using GPU programming. Especially, studying how GPGPU could be used in the scope of MABS, our main motivation is not only to speed up MABS but also to provide the MABS community with a general approach to GPU programming, which could be used on a wide variety of agent-based models. So, this paper first summarizes all the use cases that we have considered so far and then focuses on identifying which parts of the development process could be generalized.

Abstract. Using Multi-Agent Based Simulation (MABS), computing resources requirements often limit the extent to which a model could be experimented with. Regarding this issue, some research works propose to use the General-Purpose Computing on Graphics Processing Units (GPGPU) technology. GPGPU allows to use the massively parallel architecture of graphic cards to perform general-purpose computing with huge speedups. Still, GPGPU requires the underlying program to be compliant with the specific architecture of GPU devices, which is very constraining. Especially, it turns out that doing MABS using GPGPU is very challenging because converting Agent Based Models (ABM) accordingly is a very difficult task. In this context, the GPU Environmental Delegation of Agent Perceptions principle has been proposed to ease the use of GPGPU for MABS. This principle consists in making a clear separation between the agent behaviors, managed by the CPU, and environmental dynamics, handled by the GPU. For now, this principle has shown good results, but only on one single case study. In this paper, we further trial this principle by testing its feasibility and genericness on a classic ABM, namely Reynolds's boids. To this end, we first review existing boids implementations to then propose our own benchmark model. The paper then shows that applying GPU delegation not only speeds up boids simulations but also produces an ABM which is easy to understand, thanks to a clear separation of concerns.