Best practices for HPM-assisted performance engineering on modern multicore processors

摘要

Many tools and libraries employ hardware performance monitoring (HPM) onmodern processors, and using this data for performance assessment and as astarting point for code optimizations is very popular. However, such data isonly useful if it is interpreted with care, and if the right metrics are chosenfor the right purpose. We demonstrate the sensible use of hardware performancecounters in the context of a structured performance engineering approach forapplications in computational science. Typical performance patterns and theirrespective metric signatures are defined, and some of them are illustratedusing case studies. Although these generic concepts do not depend on specifictools or environments, we restrict ourselves to modern x86-based multicoreprocessors and use the likwid-perfctr tool under the Linux OS.