Performance modeling and prediction is an important and growing field. The PMaC lab at SDSC uses binary instrumentation to gather application characteristics for use in performance prediction. Our current techniques rely on using the ATOM toolkit to instrument memory references. Instrumentation overhead is a major problem. We have shown that reasonable results can be achieved using very infrequent sampling, which promises to improve instrumentation cost greatly. However, legacy tools do not adequately support such sampling. Dyninst is an excellent fit as a platform for our research goals in performance prediction. We have implemented a prototype version of our tools using Dyninst. The talk will present our experiences with that tool, what we need for our work, and then discuss some future opportunities in performance modeling and prediction that Dyninst will enable.