Simulation speed is a key issue in virtual prototyping of Multi-Processors System on Chip (MPSoCs). SystemC TLM2.0 (Transaction Level Modeling) is now commonly used to accelerate the simulation. However, the standard SystemC simulation engine uses a centralized scheduler that is clearly a bottleneck to parallelize the simulation of architectures containing hundreds of processor cores, and involving hundreds of SC_THREADs to be scheduled. In this paper, we describe a general modeling strategy for shared memory MPSoCs and associated tools for the parallel TLM simulation of these architectures. The proposed approach is based on the Parallel Discrete Event Simulation principles, and our parallel version of the SystemC kernel (named SystemC-SMP) that can run advantageously on multiprocessor workstations. As the speedup obtained by parallel simulation depends on the communication pattern between the parallel tasks, we study the influence of various locality characteristics for the software application running on the simulated MPSoC.