Measured profiling has proven to be an important tool for performance analysis of scientific applications. We believe it has significant advantages over statistical profiling methods, but there are important issues of intrusion that must be addressed. In this paper, we focus on the removal of measurement overhead from measured profiling statistics. The algorithms we describe for quantifying the overhead and eliminating it on-the-fly have been implemented in the TAU performance system. Testing these algorithms using TAU on the NAS parallel benchmarks shows that they are effective at reducing the error in estimated performance. This is demonstrated for both flat and callpath profiling. In general, the overhead compensation techniques can be applied to any set of performance metrics that can be profiled using TAU.
However, there are still concerns and problems to address. We need to validate the compensation analysis approach on other platforms where different factors will influence observed overhead. We need to better understand the limits of overhead compensation and its proper use in an integrated instrumentation and measurement strategy. In particular, the problems with overhead compensation analysis in parallel profiling require further study. While the current methods do reduce intrusion error in parallel profiling, they are unable to account for interdependent intrusion effects. We will address this problem in future research.