Once we have identified the targets (on a hybrid machine targets could include CPU and GPU nodes), create the list of applications to be measured, we must identify the experiments with the available measurements. Experiments create the set including target, application, and the performance measurement. The help “-h” following shows options for experiment creation, followed by commands creating experiments. The measurement type sample uses event based sampling for a non-instrumented executable. The profile type of measurement uses either auto-instrumentors such as PDT or compiler-instrumentation, usually invoked via a particular flag for compilers. It is possible to declare that if the auto-instrumentor fails revert back to the compiler-instrumentation.
Below we have the result of “tau dash” after the commends creating the experiments. Note that when we “tau init” we gained the “mm” application as an MPI-OpenMP application, which induced the build of the complimentary TAU that does MPI and OpenMP profiling via sampling. Listed “N/A” TAU makefiles are currently place holders, and will be populated when particular experiments are selected.