Very efficient technique for dynamically tracking locality of a reference

摘要

A cache management system determines stack distance by determining the stack reuse distance, less the number of duplicate pages in the system trace. The stack reuse distance is defined as the number of references between the first reference to a datum and the next reference to the same datum. The stack reuse distance can be easily calculated by maintaining a count of data references in a stream. The system also recognizes that the number of duplicate pages in a cache can be approximated in a probabilistic manner, thereby improving the efficiency of the cache analysis. Stack distance can be approximated during normal system operation and stack analysis can be performed in a dynamic manner for optimal cache tuning in response to data reference workloads. Thus, the cache miss ratio curve is generated either from a system trace in an off-line mode, or in real time during normal system operation.