Re: [PATCH 3/3] perf events: add timehist option to record and report

[David Ahern <daahern@cisco.com>]

On 02/18/11 11:41, Arnaldo Carvalho de Melo wrote:
> Em Fri, Feb 18, 2011 at 06:59:30PM +0100, Frederic Weisbecker escreveu:
>>   sched:sched_wait_task                      [Tracepoint event]
>>   sched:sched_wakeup                         [Tracepoint event]
>>   sched:sched_wakeup_new                     [Tracepoint event]
>>   sched:sched_switch                         [Tracepoint event]
>>   sched:sched_migrate_task                   [Tracepoint event]
>>   sched:sched_process_free                   [Tracepoint event]
>>   sched:sched_process_exit                   [Tracepoint event]
>>
>>
>> You have the sched:sched_switch event and many others.
>>
>> Just try:
>>
>> perf record -a -e sched:*
>> perf script
>>
>>             perf-4128  [000] 19242.870025: sched_stat_runtime: comm=perf pid=4128 runtime=7430405 [ns] vruntime=3530192223488 
>>             perf-4128  [000] 19242.870042: sched_stat_runtime: comm=perf pid=4128 runtime=23142 [ns] vruntime=3530192246630 [n
>>             perf-4128  [000] 19242.870045: sched_stat_sleep: comm=kondemand/0 pid=59 delay=9979163 [ns]
>>             perf-4128  [000] 19242.870048: sched_wakeup: comm=kondemand/0 pid=59 prio=120 success=1 target_cpu=000
>>             perf-4128  [000] 19242.870063: sched_stat_runtime: comm=perf pid=4128 runtime=21581 [ns] vruntime=3530192268211 [n
>>             perf-4128  [000] 19242.870066: sched_stat_wait: comm=kondemand/0 pid=59 delay=21581 [ns]
>>             perf-4128  [000] 19242.870069: sched_switch: prev_comm=perf prev_pid=4128 prev_prio=120 prev_state=R ==> next_comm
>>      kondemand/0-59    [000] 19242.870091: sched_stat_runtime: comm=kondemand/0 pid=59 runtime=27362 [ns] vruntime=35301862739
>>      kondemand/0-59    [000] 19242.870094: sched_stat_wait: comm=perf pid=4128 delay=27362 [ns]
>>      kondemand/0-59    [000] 19242.870095: sched_switch: prev_comm=kondemand/0 prev_pid=59 prev_prio=120 prev_state=S ==> next
>>
>> And you can run your own script on these events:
>>
>> $ sudo ./perf script -g python
>> generated Python script: perf-script.py
>>
>> Edit perf-script.py and then run it:
>>
>> $ perf script -s ./perf-script.py
>>
>> That also works for perl.
>>
>> The timestamps will be the cpu time and not the walltime, but at least that seems
>> to be partly what you seek?
> 
> The whole issue for him, AFAIK, is to correlate perf events with app
> events.
> 
> Think about tcpdump + networking tracepoints or 'perf probe' dynamic
> events in the network stack, he wants to merge those logs and correlate
> the tcpdump packet exchange with the tracepoints events in the network
> stack, etc.
> 
> I.e. it doesn't matter if it is ftrace or not, having a common clock
> shared between apps and kernel tracing/whatever infrastructure is what
> David is after, right?

A means of correlating the two, yes.

> 
> He can change userspace to use the clock the kernel is using in the
> perf/ftrace/whatever infrastructure or make the kernel use the clock
> userspace uses.
> 
> The issue here is who will bend, u or k ;-)

It makes sense for ftrace, perf, etc to use a monotonic clock. That
timestamp is nonsense to users (just a string of numbers) and
application writers like those pretty little strings that relate to time
has humans acknowledge it.

David

> 
> - Arnaldo