A distributed shared memory (DSM) system includes at least a first and a second nodes. The first node includes an external cache for storing a data from a local memory of the second node and at least two processors optionally accessing the data from the external cache. Whether the data has been modified into a modified data by a first certain one of the at least two processors is first determined. If positive, whether a second certain one of the at least two processors is allowed to share the modified data is further determined. If the second certain processor is allowed to share the modified data, it may directly request the modified data from the first certain processor via a bus inside the first node.