system call hardware performance counters ubuntu

Question

I am working on a project and I would like to obtain the performance counters(cache, TLB, etc) values of a system call(eg: read()) before and after the execution of a file.

I tried doing this using perf on Ubuntu but was not able to get any results. Is there a way to do it using perf or maybe some other tool ?

Thanks for the help.

3.329057 task-clock (msec)         #    0.714 CPUs utilized          
16 context-switches                #    0.005 M/sec                  
0 cpu-migrations                   #    0.000 K/sec                  
257 page-faults                    #    0.077 M/sec                  
1,983,212 cycles                   #    0.596 GHz                    
1,352,902 stalled-cycles-frontend  #   68.22% frontend cycles idle   
1,080,180 stalled-cycles-backend   #   54.47% backend  cycles idle   
1,336,919 instructions             #    0.67  insns per cycle        
                                   #    1.01  stalled cycles per insn
       267,730 branches            #   80.422 M/sec                  
       <not counted> branch-misses           

   0.004663489 seconds time elapsed

Answer 1

There is small wrapper library https://github.com/castl/easyperf for perf_event_open in counting mode, just like used by perf stat (its output was quoted by you).

You can setup hw event counting (with in kernel counting enabled - PERFMON_EVENTSEL_OS flag in easyperf), then read current values of counters perf_read_all in the wrapper, then run your function (syscall) you want to profile, and then read new counter values. Difference between old and new values is estimation of target function cost. Check this test, target function is foo:

https://github.com/castl/easyperf/blob/master/test.c

You can't measure too small calls, because reading hw counters from perf_event_open is done via several read syscalls. So, do several similar syscalls (loop of 100 or 1000), or do syscalls which does more work, or try to measure overhead of reads to get hw counters (measure empty 'foo' function to get overhead; then measure your target short function, then compare differences.)