Instrumentation in HPC++

HPC++ programs can be instrumented for tracing and replay

• Tracing
  Program execution traces can be generated for PSTL programs for a number of Event classes of Tulip. TAU provides filters for processsing these traces.
  
• Replay
  PSTL programs can be replayed using Ariadne replay facilities in TAU. It uses a deterministic replay algorithm for message passing programs using Tulip RTS and MPI. Event traces are generated using the tracing mode and the traces are merged and filtered using Ariadne filters. Replay and tracing induce an overhead on the execution that is proportional to the amount of communication..
  
• Performance profiling
  Both PSTL and HPC++lib programs can be instrumented for function and aggregate profiling.
  Function profiling shows how many times a function was invoked and the time spent in the function exclusively and inclusive of other functions that it calls. Aggregate profiling shows how many times an event took place on each aggregate spanning one or more threads.
  Currently, the user needs to insert the profiling instrumentation to identify the aggregates that should be shown. The runtime system is preinstrumented.