Synthesizing Benchmarks with Appropriate Locality and Instruction Mix

Small benchmarks that are used to measure CPU performance may not be representative of typical workloads in that they display unrealistic localities of reference and may contain unrepresentative instruction mixes. Several metrics are developed that allow more accurate development of benchmarks. The locality of reference characteristics is captured through the LRU cache hit function and the instruction mix through and execute/fetch aggregate parameterization. The operations of repetition and replication are used to form new programs and their characteristics derived. Using these techniques a general benchmark synthesis technique is obtained and demonstrated with an example.