a Q on measuring system performance.

[Peter Jeremy]

On Thu, 2005-Mar-24 23:21:54 -0800, Yan Yu wrote:
>I am trying to measure the overhead added by these instrumentation code.
>my plan is:
>  in my user space program, i have something like the following:
>--------------------------------------------
>        gettimeofday(&prev_time, NULL);
>        for (i=0; i< 1000; i++)
>        {
>                fd = fopen("tmp", "r" );
>                if (fd == NULL)
>                {
>                        break;
>                }
>                cnt ++;
>        }
>
>        gettimeofday(&cur_time, NULL);
>        t_lapse= misc_tv_offset( &cur_time, &prev_time );
>
>----------------------------------------------------

That approach is reasonable (but the above code leaks file
descriptors) .  You might want to increase 1000 and/or use rdtsc() for
timing depending on your accuracy/resolution requirements.

>I would run this for the unmodified kernel, and instrumented kernel.
>compare the t_lapse, my concern is that t_lapse includes context switch
>time when the user process is taken out of run queue.

So would gprof.  And gprof has much higher overheads and a granularity
of 10usec.

-- 
Peter Jeremy