Our research focuses on the location and explanation of performance problems in parallel programs, a task that we call performance diagnosis. Researchers have developed many software tools to collect and analyze data for performance diagnosis, but many obstacles prevent such tools from practically benefiting parallel programmers. Two obstacles in particular motivate our current work:

1. Researchers lack a theory of what methods work, and why. There is no formal way IO describe or compare the ways expert programmers solve their performance diagnosis problems in particular contexts.
2. Parallel programmers lack a guide to what tools work, and where. There is no standard framework for understanding tool features and fitting them to the programmer's particular needs.

Both of these obstacles, we believe, could be mitigated by a formal theory of performance diag­nosis processes. This article summarizes such a theory, and proposes a research collaboration to evaluate that theory against actual programming practices.