PerformanceProfiling

PFunc is fully integrated with the Performance Application Programming Interface (PAPI), thereby allowing users to profile their applications with ease. PAPI was chosen due to its wide availability and portability across many hardware platforms. Profiling is handled mainly through the taskmgr type. Users can specify the events (both PAPI presets and native) that they wish to monitor when initializing objects of the taskmgr type. Consider the sample code given below:

typedef pfunc::generator<cilkS, /* scheduling policy */
                         pfunc::use_default, /* compare */
                         pfunc::use_default> my_pfunc; /*function object*/

enum {nthds=2,
      nevents=2
};

struct my_perf_data: pfunc::perf_data {
  perf_data::event_value_type storage[nthds][nevents];
  int events[nevents];

  my_perf_data () {
    events[0] = PAPI_L1_DCA; 
    events[1]= PAPI_L1_DCM;
  }

  int get_num_events () const { return nevents; }

  int* get_events () const { return events; }

  perf_data::event_value_type** get_event_storage () const {
    return storage; 
  }
};

my_perf_data perf;
my_pfunc::taskmgr cilk_tmanager (/*nqueues*/, /*nthreads_per_queue*/, perf);

In this example, we utilize PFunc to measure L1 data cache behavior. We derive from perf_data to communicate the required measurements to PFunc. PFunc stores the requested event values in my_perf_data and these values can subsequently be used for performance tuning.