Connection through a single bus is the simplest architecture for multiprocessor systems. However, as the number of processors scales up, the common bus will eventually become the bottleneck of the system. Three different bus‐based architectures, multiprocessors without local memories, multiprocessors with single port local memories, and multiprocessors with dual port memories, are analyzed in this paper. It shows to what degree increasing the local execution probability of a multiprocessor system with dual port memories can relieve the load of the common bus and thus increase the effective computing power. A multiprocessor operating system (MOS) designed at National Chiao Tung University (NCTU) has suffered bottleneck trouble in the previous year. Problems in MOS are discussed. A new design of this operating system kernel, MOS*, is then proposed. Performance of these two different designs are evaluated and compared by simulation. The simulation results show that MOS*, with a localizing process strategy, can increase effective computing power 30–100 percent. The saturation point of MOS* is also explored in the paper.