Score-P metrics
Time
- Description:
-
Total time spent for program execution including the idle times of CPUs
reserved for slave threads during OpenMP sequential execution. This
pattern assumes that every thread of a process allocated a separate CPU
during the entire runtime of the process.
- Unit:
- Seconds
- Diagnosis:
-
Expand the call tree to identify important callpaths and routines where
most time is spent, and examine the times for each process or thread to
locate load imbalance.
Visits
- Description:
-
Number of times a call path has been visited. Visit counts for MPI routine
call paths directly relate to the number of MPI communications and
synchronizations. Visit counts for OpenMP operations and parallel regions
(loops) directly relate to the number of times they were executed.
Routines which were not instrumented, or were filtered during measurement,
do not appear on recorded call paths. Similarly, routines are not shown if
the compiler optimizer successfully in-lined them prior to automatic
instrumentation.
- Unit:
- Counts
- Diagnosis:
-
Call paths that are frequently visited (and thereby have high exclusive
Visit counts) can be expected to have an important role in application
execution performance (e.g., Time). Very frequently
executed routines, which are relatively short and quick to execute, may
have an adverse impact on measurement quality. This can be due to
instrumentation preventing in-lining and other compiler optimizations
and/or overheads associated with measurement such as reading timers and
hardware counters on routine entry and exit. When such routines consist
solely of local/sequential computation (i.e., neither communication nor
synchronization), they should be eliminated to improve the quality of the
parallel measurement and analysis. One approach is to specify the names of
such routines in a filter file for subsequent measurements to
ignore, and thereby considerably reduce their measurement impact.
Alternatively, selective instrumentation can be employed to
entirely avoid instrumenting such routines and thereby remove all
measurement impact. In both cases, uninstrumented and filtered routines
will not appear in the measurement and analysis, much as if they had been
"in-lined" into their calling routine.
Hits
- Description:
-
Number of exclusive samples inside this region.
- Unit:
- Counts
Wrapped libraries
- Description:
-
Total time spent for program execution in external libraries.
- Unit:
- Seconds