`perf` instructions incomplete

[arnetheduck]

To measure wall clock time, perf -e task-clock can be used:

perf record --call-graph dwarf,16000 -e task-clock ./testProf

    99.92%     0.19%  testProf  testProf                   [.] main
            |          
             --99.73%--main
                       readRandFileValue
                       |          
                       |--96.41%--__GI_fseek (inlined)
                       |          |          
                       |           --95.45%--_IO_new_file_seekoff (inlined)
                       |                     |          
                       |                     |--46.41%--__GI___libc_read (inlined)
                       |                     |          
                       |                      --45.47%--__lseek64 (inlined)
                       |          
                        --2.41%--getRandomBoundedInt
                                  |          
                                   --1.15%--rand
                                             |          
                                              --1.05%--__random
                                                        |          
                                                         --0.73%--__random_r

[jasonrohrer]

Hey, that's very cool.

I guess perf works to do this after all.

I will leave this here for others to see. I may update the docs at some point in the future, but it's not a high priority right now.

[jasonrohrer]

Actually, where is this documented? Not in the man page:

https://man7.org/linux/man-pages/man1/perf.1.html

Maybe it's a newer feature? Even the -e switch isn't mentioned there.

[arnetheduck]

-e is documented in the more specific pages like:

https://man7.org/linux/man-pages/man1/perf-record.1.html https://man7.org/linux/man-pages/man1/perf-stat.1.html#EXAMPLES

It takes a bit of digging though, in particular to figure out what task-clock actually means: https://stackoverflow.com/questions/23965363/linux-perf-events-cpu-clock-and-task-clock-what-is-the-difference

[lieff]

task-clock does`t work for me. Shows only task execution time, not wait of any kind. Quick check:

$ perf stat -e task-clock sleep 1                                                                                                                                                                                                      ↑

 Performance counter stats for 'sleep 1':

              0,53 msec task-clock                #    0,001 CPUs utilized

       1,001286900 seconds time elapsed

       0,001180000 seconds user
       0,000000000 seconds sys

no matter how many seconds to wait.

It's just pure luck that __GI_fseek execution time above anything too in report above. But it does`t really count wait time.