An Adaptive Cost Model for Parallel Program Instrumentation

Software based instrumentation is frequently used to measure the performance of parallel and distributed programs. However, using software instrumentation can introduce serious perturbation of the program being measured. In this paper we present a new data collection cost system that pro-vides programmers with feedback about the impact data collection is having on their application. In addition, we introduce a technique that permits programmers to define the perturbation their appli-cation can tolerate and then we are able to regulate the amount of instrumentation to ensure that threshold is not exceeded. We also describe an implementation of the cost model and presents results from using it to measure the instrumentation overhead for several real applications. 1