The interconnection network’s design of large-scale parallel computer systems is significant to the execturion of the efficiency of parallel programs. Currently, the Petaflop supercomputers usually have more than ten thousand computing nodes, which cause new challenges to the performance of interconnection network. However, most exited studies only consider simple network workload models which have many differences from the workloads of real parallel applications. This paper presents complex workload models which are more similar to the practical network workload. Then based on the mathematical model of the interconnection network in the earlier study, the study uses a flit-level network simulator to support analyzing those complex workload models. Finally, through a great deal of experiments the performance of Torus and Fat Tree network topologies are compared under different workload models. Meanwhile, the message mean latency of the 3D FFT parallel algorithm using 2D partitioning is simulated. The results and related analysis efficiently support the large-scale interconnection network’s design, as well as the optimization of parallel programs.