Using KUPL PROF for Full-Scale Function Timeline Profiling

KUPL provides two tools, profile and trace, to provide runtime timeline analysis. The profile and trace capabilities are enabled by dynamically linking to the corresponding prof shared library.

1. Install HPCkit by following the instructions in Kunpeng HPCKit 26.1.RC1 Installation Guide.
2. Run the setvars.sh --kupl_mode=prof command to link to the prof shared library.
3. When running the program, configure the environment variable KUPL_PROF_LEVEL=statistic to enable cumulative time profiling for specific functions, and configure the environment variable KUPL_PROF_LEVEL=trace to enable full-scale function timeline tracing.
4. Upon application completion, if KUPL_PROF_LEVEL is set to either statistic or trace, a cumulative time profile report for functions will be printed at the end of the execution output (see Table 1 Header fields in the cumulative time profile report for functions for details). If KUPL_PROF_LEVEL=trace is specified, a ptrace file (in JSON format) capturing the per-process and per-thread execution traces will be generated in the application's working directory. This file can be imported into the chrome://tracing UI (Chrome 94 or later is required) for visual timeline analysis.

   Table 1 Header fields in the cumulative time profile report for functions

   Field	Description
   thread_id	Execution thread ID
   msg_id	Function event number
   msg	Function event name
   sum(ms)	Total function duration (in milliseconds)
   count	Total number of function executions
   avg(ns)	Average function duration (in nanoseconds)
   max(ns)	Maximum execution time for a single function call (in nanoseconds)
   min(ns)	Minimum execution time for a single function call (in nanoseconds)

   Table 2 msg fields in the cumulative time profile report for functions

   msg Field	Function Behavior
   task_execute	Duration of executing tasks submitted to KUPL
   task_finish	Duration of releasing resources after tasks submitted to KUPL complete execution
   sched_add_taskbase	Duration of submitting tasks to the underlying KUPL scheduler (Sched)
   sched_get_taskbase	Duration of retrieving tasks from the underlying KUPL Sched

   

   It is advised to use Google Chrome to view the visualized report.