Abstract
In recent years, software infrastructures for the run-time instrumentation of programs have begun to emerge. This paper presents and evaluates prototypes of dynamic instrumentation and dynamic control of instrumentation for parallel mixed MPI/OpenMP applications. An overview of the technology behind these approaches is presented. Prototypes of dynamic instrumentation and dynamic control of instrumentation for use with the Vampir/GuideView (VGV) toolset are discussed. Instrumentation evaluations using the ASCI kernel benchmarks are used for proof of concept on a cluster of SMPs. The results demonstrate that a mix of dynamic instrumentation and dynamic control of instrumentation can be an effective performance analysis alternative to the traditional static instrumentation of applications.