Messages in this thread | | | Date | Mon, 2 Nov 2015 20:04:36 -0300 | From | Arnaldo Carvalho de Melo <> | Subject | Re: [RFC/PATCH 0/4] perf report: Support folded callchain output (v2) |
| |
Em Tue, Nov 03, 2015 at 07:49:27AM +0900, Namhyung Kim escreveu: > On Mon, Nov 02, 2015 at 07:28:42PM -0300, Arnaldo Carvalho de Melo wrote: > > Hi Namhyung, > > > > Em Tue, Nov 03, 2015 at 07:12:04AM +0900, Namhyung Kim escreveu: > > > On Mon, Nov 02, 2015 at 06:30:21PM -0300, Arnaldo Carvalho de Melo wrote: > > > > Em Mon, Nov 02, 2015 at 12:37:28PM -0800, Brendan Gregg escreveu: > > > > > On Mon, Nov 2, 2015 at 4:57 AM, Namhyung Kim <namhyung@kernel.org> wrote: > > > > > > This is what Brendan requested on the perf-users mailing list [1] to > > > > > > support FlameGraphs [2] more efficiently. This patchset adds a few > > > > > > more callchain options to adjust the output for it. > > > > > > > > At first, 'folded' output mode was added. The folded output puts all > > > > > > calchain nodes in a line separated by semicolons, a space and the > > > > > > value. Now it only supports --stdio as other UI provides some way of > > > > > > folding/expanding callchains dynamically. > > > > > > > > The value is now can be one of 'percent', 'period', or 'count'. The > > > > > > percent is current default output and the period is the raw number of > > > > > > sample periods. The count is the number of samples for each callchain. > > > > > > > > Here's an example: > > > > > > > > $ perf report --no-children --show-nr-samples --stdio -g folded,count > > > > > > ... > > > > > > 39.93% 80 swapper [kernel.vmlinux] [k] intel_idel > > > > > > intel_idle;cpuidle_enter_state;cpuidle_enter;call_cpuidle;cpu_startup_entry;start_secondary 57 > > > > > > intel_idle;cpuidle_enter_state;cpuidle_enter;call_cpuidle;cpu_startup_entry;rest_init;... 23 > > > > > > > So for the folded output I don't need the summary line (the row of > > > > > columns printed by hist_entry__snprintf()), and don't need anything > > > > > except folded stacks and the counts. If working with the existing > > > > > stdio interface is making it harder than it needs to be, might it be > > > > > > I don't think it so, just add some flag asking for that > > > > hist_entry__snprintf() to be supressed, ideas for a long option name? > > > > > > Having it as Namhyung did may have value for some people as a more > > > > compact way to show the callchains together with the hist_entry line. > > > > > Yeah, I'd keep the hist entry line unless it's too hard to > > > parse/filter. IMHO it's just a way to show callchains, so no need to > > > > What I suggested was to have something like: > > > > $ perf report --no-children --no-hists --stdio -g folded,count > > ^^^^^^^^^^ > > ^^^^^^^^^^ > > ... > > intel_idle;cpuidle_enter_state;cpuidle_enter;call_cpuidle;cpu_startup_entry;start_secondary 57 > > intel_idle;cpuidle_enter_state;cpuidle_enter;call_cpuidle;cpu_startup_entry;rest_init;... 23 > > > > I.e. the first entry in the callchain is 'intel_idle', just like in what > > Brendan called the 'summary line', i.e. reduntant when what he wants its > > just all the callchains and how many times they were sampled. > > Yep, I know. But isn't 'perf report' all for seeing hist lines? :)
Well, so far, yes, but he is presenting a usecase where what we want to see is just callchains, and we can achieve that rather easily, no?
> I'm not insisting it strongly, but it's a bit strange for me if perf > report doesn't show any hist lines..
If that is of no use in this use case, why not?
> > > have separate output mode.. > > > > > Brendan, I guess you still need to know other info like cpu or pid, no? > > > > Possibly, but just with the callchains he has enough info for the basic > > flame graph, no? > > > > > And I feel like it'd be better to put the count before the callchains > > > for consistency like below. Is it OK to you? > > > > Consistency with what? > > Oh, I meant consistency with other callchain output style like graph, > fractal or flat - They all show the numbers before callchains. And I > think it's easier to read for human. :)
Well, As I said, isn't the main object here the callchain? :-)
And Brendan's request is for a something to be consumed by scripts, i.e. something like we have for perf stat:
For humans:
[root@felicio ~]# perf stat -e cycles -I 1000 -a # time counts unit events 1.000304391 1,820,038 cycles 2.000490191 1,005,477,007 cycles 3.000657813 1,717,007 cycles ^C 3.917890293 2,804,034 cycles
For machines/scripts:
[root@felicio ~]# perf stat -x, -e cycles -I 1000 -a 1.000291954,1923360,,cycles,3998167210,100.00 2.000477154,1005608105,,cycles,3998475482,100.00 3.000612612,1345483,,cycles,3998332391,100.00 4.000744469,1005046913,,cycles,3998258199,100.00 ^C 4.331684347,1551327,,cycles,3463190970,100.00
[root@felicio ~]#
> > The main thing here is the callchain, all the other stuff are things > > related to it, so showing it first makes sense to me. > > > > Having some way to list the desired info to have for each callchain may > > be interesting, and if he could do it like: > > > > -g folded,count,cpu,other,fields > > > > then he would know how to parse the per-callchain info at the end of > > each line, right? > > Hmm.. looks like that it ends up having redundant info. I don't think
What is redundant, and with with what?
> it's generally useful to other 'perf report' stuffs. Wouldn't it be > better just adding minimal support and let the external tool parse the > output?
Oh well, perhaps we could have a 'perf callchain' tool that would be centered on callchains and would provided one line per callchain, which would have:
callchain;seprarated;colons series,of,desired,fields,for,this,callchain
Which would reuse heavily the 'perf report' / 'perf top' code for histograms, no?
I still think that this is a 'perf report' thing, but one that is centered in callchains, and that is to be consumed by scripts, not humans.
- Arnaldo
| |