Parallel computing is currently the only possible method for providing sufficient performance of large scale urban traffic simulations. The need for representing large areas with detailed, continuous space and motion models exceeds the capabilities of a single computer in terms of performance and memory capacity. Efficient distribution of such computation, which is considered in this paper, poses a challenge due to the need of repetitive synchronization of the common simulation model. We propose an architecture for efficient memory and communication management, which allows executing simulations of huge urban areas and efficient utilization of hundreds of computing nodes. In addition to analyzing performance tests, we also provide general guidelines for designing large-scale distributed simulations.