This thesis aims to design a distributed large scale MAS simulation. When the number of agents reaches several millions, it is necessary to distribute MAS simulation. However, this can raise some issues: agents allocation, interactions from different machines, time management, etc. When we distribute MAS simulation on different machines, agents must be separated between these machines and should still be able to produce their normal behaviours. Our distribution is able to cover all agents' perceptions during the simulation and allow all agents to interact normally. Moreover, with large-scale simulations the main observations are done on the macroscopic level. In this thesis, we study two main aspects to distribute large-scale simulations. The first aspect is the efficient strategy that can be used to distribute MAS concepts (agents and environment). We propose two efficient distribution approaches: agents distribution and environment distribution. The second aspect is the relaxation of synchronization constraints in order to speed up the execution of large-scale simulations. Relaxing this constraint can induce incoherent interactions, which do not exist in a synchronized context. But, in some applications that can not affect the macroscopic level. Our experiments on different categories of MAS applications show that some applications can be distributed efficiently in one distribution approach more than the other. In addition, we have studied the impact of incoherent iterations on the emerging behaviour of different applications, and we have evidenced situations in which unsynchronized simulations still produced the expected macroscopic behaviour.