The success of the TAU performance system thus far has been primarily due to the benefits of targeting the TAU performance technology to the general computation model and integrating the technology within the framework architecture. TAU functionality is governed, therefore, primarily by how performance observation requirements are addressed at an abstract computation level, allowing system-dependent implementation to be concerned with how efficiently the observation of abstract performance events is supported for a specific platform. The outcome has been broad availability of TAU technology across:
In addition, TAU has been able to integrate effectively technology from other projects: DyninstAPI [3], PCL [14], PAPI [23], Vampir [12], and PDT [22]. The successes of TAU also represent its opportunity for enhancement. TAU's architecture will likely need extending, especially to better support dynamic performance measurement and adaptive performance control. Plus, the general computational model may need further sophistication. Nevertheless, we believe the TAU system as a whole captures general observational and analysis techniques and components that are foundational in general performance technology evolution.