Abstract

Distributed shared memory (DSM) architectures try to bind recently referenced data to memory local to a node with the assumption that the data will be referenced again in the near future. However, patterns of reference at the local memory may not be as highly localized as at the processor cache. In general, the higher the cache hit rate, the lower the local memory hit rate. Also, unlike in traditional caches, replaced data from the local memory can have the side effect of affecting locality of data in a remote memory. Thus, utilizing locality of reference at the local memory seems to be arduous in DSM architectures due to higher overhead for dynamic binding and weaker locality of reference. This thesis considers three binding times for new data at the local memory: page fault time, node miss time, and cache miss time. In DSM architectures, different binding times are associated with different ways of arranging data, which cause different overheads under the same patterns of reference. ...

Citations