Compiler-Based Approaches to Reduce Memory. Access Penalties in Cache Coherent Multiprocessors

Doctoral thesis, 1997

To reduce the average time needed to perform a read or a write access in a multiprocessor, a cache is associated with each processor. A hardware mechanism is used to ensure that the replicated cache copies are consistent. This mechanism employs a protocol which controls when a node may read and/or write a shared data item. The time a processor is waiting for actions of the protocol to be performed, called the memory access penalty, limits the performance that can be achieved.