In high-performance computing of multi-agent systems, there often exists a load imbalance that slows down the calculation. In this paper, we discuss the parallelization of the Boids model for simulating a swarm intelligence. We apply the multi-level hierarchy of parallelism to the Boids model to mitigate the load-imbalance problem. To eliminate numerical errors due to parallelization, we apply pseudo-quadruple arithmetic. The parallel performance is evaluated on three major architectures, including many-core processors on an x86-based server with GPUs, and the Earth Simulator. The parallelization can decrease the negative effects of a load imbalance to almost zero in a simulation of 50 million agents. In addition, the parallelization guarantees the reproducibility of the results in a sequential execution. The strong scaling shows the potential to complete a simulation in real-time on the Earth Simulator. In addition, the weak scaling shows the ability to calculate 100 billion agents within a reasonable amount of time.