Memory Block Relocation in Cache-Only Memory Multiprocessors.
Gyungho Lee, Sujat Jamil · 1995
Designed to provide a high utilization of local memory in each processing node, the cache-only memory architecture (COMA) is a type of distributed shared memory multiprocessor, where the local memory serves as a large cache, called attraction memory (AM), to the shared address space. When a memory reference misses in the local AM, the relevant memory block is copied from a remote AM to the local AM, replacing a memory block from the local AM if necessary, due to lack of space in the corresponding AM set. Since main memory in the traditional sense is absent in this architecture, there is no backing store in primary storage to write back a replaced memory block. Also, writing a replaced block back to secondary storage incurs unacceptable overhead. Therefore, a replaced block with no copies in other AMs need to be relocated to another AM. The simplest solution is to relocate a replaced block to the node that supplied the block satisfying the miss. However, we show in this paper that this ...