Profile-based optimization has been studied extensively. Numerous papers and real systems have shown substantial improvements. However, most of these papers have been limited to either branch prediction or instruction cache performance. Also, most of these papers have looked at small applications with a limited number of testing and training scenarios.

In this paper, we look at real use of large real-world desktop applications. We also assume memory consumption and disk performance are the primary metrics of interest. For this domain, we show that it is very difficult to get adequate coverage of large applications even with an extensive collection of training scenarios. We propose instead to augment traditional scenarios with data derived from real use. We show that this methodology allows us to reduce memory pressure by 29% and disk reads by 33% compared to traditional approaches.