Given the TAU performance measurement technology, the important question becomes what is the approach best suited for component performance measurement. There are two measurement types we envision based on how a component is instrumented: 1) with direct calls to a measurement library or 2) using an abstract measurement interface. The difference is depicted in Figure 4. TAU specializes in multi-level performance instrumentation targeting a common performance measurement API. We can instrument the component code to call TAU measurement routines directly using the API, as shown in the left part of Figure 4. To facilitate the instrumentation, TAU provides automated source instrumentation tools (based on the Program Database Toolkit (PDT) [24]) and dynamic instrumentation support. As shown, the TAU measurement system maintains runtime performance data that can be accessed via the API directly or stored in files at the end of component execution.
In contrast, the component could also be instrumented to call an abstract measurement component interface, as shown in the right part of Figure 4. This interface would be implemented by a performance component that targets a backend measurement system (in this case TAU). There are several benefits to this approach. First, a component could be developed with ``virtual instrumentation'' in the sense that the abstract measurement interface is virtual (i.e., consists of virtual functions). The overhead of instrumentation is nullified until a performance component is instantiated. Second, it is possible to use any measurement system in the performance component that can conform to the interface. Lastly, the performance component can provide ports for other components to use, including ports to access performance data without touching the instrumented application component. This raises the possibility that the application component is instrumented directly, but the performance data is accessed via the performance component. The downside of this approach is that the measurement interface is possibly less efficient or that it does not allow certain types of detailed performance measurements to be made.
