This work explores the possibility of using speculation at the directories in a cache coherent non-uniform memory access multiprocessor architecture to improve performance by forwarding data to their destinations before requests are sent. It improves on previous consumer prediction techniques, showing how to construct a predictor that can handle a tradeoff of accuracy and coverage. This dissertation then explores the correct time to perform consumer prediction, and show how a directory protocol can incorporate such a scheme. The consumer prediction enhanced protocol that is developed is able to reduce the runtime of a set of scientific benchmarks by 10%-20%, without substantially reducing the runtime of other benchmarks; specifically, those benchmarks feature simple phased behavior and regularly distribute data to more than two processors.

This work then explores the interaction of consumer prediction with two other forms of prediction, migratory prediction and last touch prediction. It demonstrates a mechanism by which migratory prediction can be implemented using only the storage elements already present in a consumer predictor. By combining this migratory predictor with a consumer predictor, it is possible to produce greater speedups than did either individually. Finally, the signatures of the last touch predictor can be applied to improve the performance of consumer prediction.