A framework that automatically generates synthetic benchmarks for multicore systems is presented in this paper.

A synthetic benchmark emulates a provided application in terms of its characteristics, for example, instruction count by performing (artificial) operations repeatedly in a loop. The goals are to reduce an application’s runtime and program code size; both are beneficial for benchmarking hardware simulators since they are typically slower than real machines by orders of magnitude.

The benchmarks are generated in the (high-level) programming language C, which makes the program code portable as well as human-readable (in contrast to low-level assembly language, which is generated by most of the related work). It could be shown that synthetic benchmarks are on average 147 times faster and 11 times smaller than emulated (real-life) benchmarks of the common suites PARSEC and Rodinia.