How TAU works in C++ code (theory of operation)

To profile C++ code, we need to instrument the source code using the TAU Profiling API. Instrumenting the sources involves identifying each function uniquely and associating it with one or more Profile groups.

• Profiling Templates
  Instantiations of templated entities uniquely identify the functions

  template <class T>
  class TemplClass { // class template with Get and Set methods
    private :
      T Data;
    public :
  
    T SetData(T d) {
      Data = d;
      return Data;
    }
  }
  
  

  What is the time spent in ?

  TemplClass::SetData()  int(int)
  TemplClass::SetData()  vector<int> (vector<int>) 
  TemplClass::SetData()  double(double)
  
  

  Adding TAU Profiling API calls to your code
  Each Function is uniquely identified by

  1. Unique Fuction name e.g., "TemplClass::SetData()"
  2. Type information that includes the return type and the type information of arguments. e.g., "int (int)", "double (double)"

  Different instantiations of templated entities can be identified at runtime using the macro CT(obj). It returns the string containing the runtime type information of that object. e.g. for the above example :

   TemplClass<float> obj1; 
       in TemplClass::SetData (T d) 
       CT(d) returns "float"
  

  The profiled code looks like :

  template <class T>
  class TemplClass { // class template with Get and Set methods
    private :
      T Data;
    public :
  
    T SetData(T d) {
      TAU_TYPE_STRING(taustr, CT(Data) + " (" + CT(d) + " )" );
      TAU_PROFILE("TemplClass::SetData()", taustr, TAU_DEFAULT);
      Data = d;
      return Data;
    }
  }
  

  Adding instrumentation is covered in greater detail in the tutorial .

  Terminology : Exclusive and Inclusive Time

   
  int f1() 
  { // 100 secs total time inclusive of others
     f2(); // 10 secs
  
  ...
     f3(); // 10 secs 
  
  } // 80 secs exclusively spent in f1()
  

  Exclusive time referes to the time spent in a function exclusive of the time spent in all profiled functions that it calls. Inclusive time is the total time taken in any given function.

  Hardware Counters
  We can replace "time" by hardware counters and can get exclusive and inclusive counts for each function. The counters could refer to no. of floating point operations , cycles , instructions issued, secondary data cache misses etc. This involves recompilation with a different set of switches.

  Selective Profiling
  Each function is associated with one or more profile groups. While runing the profiled application the user can specify which profile groups are to be enabled (by default all are enabled) and so we can selectively profile a group of functions.

  Runtime Switches
  The user can specify runtime switches for selectively profiling a subset of profiled functions on the command line. The list of switches is available at here .

  By default the profile data files are generated for each node/context/thread in the current directory (named profile.<node>.<context>.<thread>). To store the data in some other directory, the environment variable PROFILEDIR should be set and should point to the appropriate location.

  If we're profiling using hardware counters instead of time we need to set an environment variable T5_EVENT0 on the SGI R10000 platform. The list of legal values that it can take and the parameter that it represents can be found here .

  Using pprof and racy
  After running the program and generating the profile data files, the user can invoke pprof - the parallel profile data display tool which shows the text output or racy the graphical user interface to pprof.