Due to the ease of implementation and the alleviation of memory bottleneck effect, NUMA system with Remote Access Cache (RAC) has dominated multiprocessor systems for the past several years. In this paper, we suggest a Multi-Grain RAC to adaptively control the RAC line size for each application behavior. We simulate the NUMA systems with the Multi-Grain RAC using MINT, an event-driven memory hierarchy simulator, and analyze the performance of Multi-Grain RAC. At first, with a profile- based determination method, we verify the best RAC line size for each application and then we compare and analyze the performance differences among NUMA systems with difference RAC line size. The simulation result shows that the worst case can be mostly avoided in Multi-Grain RAC and, in addition, the results are very close to the best case with any combination of application and RAC format.