Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1751519AbaDPI3T (ORCPT ); Wed, 16 Apr 2014 04:29:19 -0400 Received: from LGEMRELSE6Q.lge.com ([156.147.1.121]:42431 "EHLO lgemrelse6q.lge.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751007AbaDPI3N (ORCPT ); Wed, 16 Apr 2014 04:29:13 -0400 X-Original-SENDERIP: 10.177.220.181 X-Original-MAILFROM: namhyung@gmail.com From: Namhyung Kim To: Don Zickus Cc: acme@kernel.org, jolsa@redhat.com, eranian@google.com, Andi Kleen , LKML Subject: Re: [RFC 0/5] perf: Create hist_entry groups References: <1397160661-33395-1-git-send-email-dzickus@redhat.com> <878ur7thwh.fsf@sejong.aot.lge.com> <20140415160841.GT8488@redhat.com> Date: Wed, 16 Apr 2014 17:29:09 +0900 In-Reply-To: <20140415160841.GT8488@redhat.com> (Don Zickus's message of "Tue, 15 Apr 2014 12:08:41 -0400") Message-ID: <87sipdsmne.fsf@sejong.aot.lge.com> User-Agent: Gnus/5.13 (Gnus v5.13) Emacs/24.1 (gnu/linux) MIME-Version: 1.0 Content-Type: text/plain Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Hi Don, On Tue, 15 Apr 2014 12:08:41 -0400, Don Zickus wrote: > On Tue, Apr 15, 2014 at 12:01:50PM +0900, Namhyung Kim wrote: >> Hi Don, >> >> On Thu, 10 Apr 2014 16:10:56 -0400, Don Zickus wrote: >> > This patchset creates a new layer of hist entry objects called >> > hist_entry_groups. The purpose is to help organize the hist_entries >> > into groups before sorting them. As a result you can gain a >> > new perspective on the data by organizing the groups into cpu, pid >> > or cacheline. See patch 5 for sample output. >> > >> > The main driver for this patchset is to find a way to sort and display >> > cacheline data in a way that is useful. My previous attempts seemed >> > hackish until I realized cacheline sorting is really just a collection >> > of hist_entries. Anyway that was my focus for doing this. >> > >> > The overall idea looks like: >> > >> > evlist >> > evsel >> > hists >> > hist_entry_group <<< new object >> > hist_entry >> > >> > >> > Implementing this was not pretty. I tried to seperate the patches the >> > best I could. But in order for each patch to compile, patch 4 turned into >> > a 1400 line diff that is mostly noise. >> > >> > Also, this patchset breaks most tools (mainly because I don't understand >> > all the interactions), hence the RFC. I mostly tested with 'perf report >> > --stdio' and 'perf mem report --stdio'. >> > >> > Please let me know if this is an interesting idea to go forward with or not. >> >> I'd like to show you my previous two patchsets. >> >> The first one is for adding --field option and changing the sort >> behavior little different [1]. I'm about to send a new version to the >> list soon. >> >> I think what you want to do is sorting output by an order of sort keys >> not just by the overhead. So with the patchset applied, you can do it >> like: >> >> $ perf report --field overhead,pid,dso,sym --sort pid >> >> # Overhead Command: Pid Shared Object >> # ........ .................... ................. ........................... >> # >> 32.93% swapper: 0 [kernel.kallsyms] [k] intel_idle >> 6.79% swapper: 0 [kernel.kallsyms] [k] enqueue_entity >> 1.42% swapper: 0 [kernel.kallsyms] [k] update_sd_lb_stats >> 1.30% swapper: 0 [kernel.kallsyms] [k] timekeeping_max_deferme >> 1.18% swapper: 0 [kernel.kallsyms] [k] update_cfs_shares >> 1.07% swapper: 0 [kernel.kallsyms] [k] __irq_work_run >> 0.96% swapper: 0 [kernel.kallsyms] [k] rcu_check_callbacks >> 0.64% swapper: 0 [kernel.kallsyms] [k] irqtime_account_process >> 0.50% swapper: 0 [kernel.kallsyms] [k] int_sqrt >> 0.47% swapper: 0 [kernel.kallsyms] [k] __tick_nohz_idle_enter >> 0.47% swapper: 0 [kernel.kallsyms] [k] menu_select >> 0.35% swapper: 0 [kernel.kallsyms] [k] run_timer_softirq >> 0.16% swapper: 0 [kernel.kallsyms] [k] __perf_event_enable >> 0.12% swapper: 0 [kernel.kallsyms] [k] rcu_eqs_exit_common.isr >> 0.50% watchdog/6: 37 [kernel.kallsyms] [k] update_sd_lb_stats >> 3.45% Xorg: 1335 [kernel.kallsyms] [k] schedule >> 6.55% gnome-terminal: 1903 libc-2.17.so [.] __strcmp_sse42 >> 1.59% firefox: 2137 [kernel.kallsyms] [k] cpuacct_charge >> 0.50% emacs: 2473 emacs-24.1 [.] 0x000000000012241a >> 0.38% emacs: 2473 emacs-24.1 [.] 0x00000000000bfbf7 >> 0.31% emacs: 2473 emacs-24.1 [.] 0x00000000001780dd >> 0.29% emacs: 2473 emacs-24.1 [.] 0x000000000002eb48 >> 4.40% kworker/7:1:11028 [kernel.kallsyms] [k] generic_exec_single >> 1.30% kworker/0:0:25667 [kernel.kallsyms] [k] generic_exec_single >> 5.93% kworker/5:1:26447 [kernel.kallsyms] [k] generic_exec_single >> 2.06% kworker/1:2:26653 [kernel.kallsyms] [k] generic_exec_single >> >> As you can see the output is now sorted by pid value (and then overhead, >> dso, sym if previous key resulted in a same value), so swapper (pid 0) >> comes first and then watchdog/6, Xorg, and so on.. > > This is probably a workable solution for our c2c tool. I can play with > this some more. Cool. :) > >> >> But it's not guarantee that the hottest pid comes always first on the >> output, it just sorted it by pid and it gets the result simply because >> the system was idle mostly. I think you can handle it in your c2c tool >> properly though. >> >> Another one I'd like to introduce is somewhat similar to your work. >> It's called hierarchy view and groups each entries according to sort >> keys [2]. But it only supported --gtk output at that time (in order not >> to make the hands dirty unnecessarily ;-) and (thus?) didn't get much >> review. But I think the idea is same and requires less change by just >> adding few fields (rb_root) to hist_entry instead of new data structure. > > Looks promising. > > I keep thinking with all these hist_entry hacks to support flexibility, if > we should just do some bigger changes to the design. I was thinking along > the lines of combining hist_entries and callchain stuff and maybe output > changes into a unified heirarchy somehow. This way we could re-use alot > of code and throw away all the silly callchain special cases and just > treat it like a sort_entry. > > I am not sure how that would work (or if really possible), but I was > playing with ideas in my head based on Jiri's suggestion, of something > like a tree layout where 'struct hists' would be sorta like a directory > and would dictate the data type in the 'files' of 'struct hist_entry'. > > The idea was 'struct hists' would normally have a HIST data type and > contain the specific sort_entry(ies) for its heirarchy. The 'struct > hist_entries' below it would all be the normal HIST data type. For > callchain support, there would be a 'struct hist' under each 'hist_entry' > that would be of data type CALLCHAIN and its sort specific rules. > > This way we could add display a callchain anywhere in the heirarchy > (instead of the normal last position). I don't understand what you want to do - having callchains in the middle is not intuitive to me. Btw, you may want to check my cumulative (or children) work which adds callchains into normal output. https://lkml.org/lkml/2014/3/20/50 > > If you then split the entries and entries_in out of struct hist and > instead create two 'struct hists', one for input and one output. Then > perhaps we could create a data type GTK_OUT for a gtk specific output sort > of entries. This might help re-use/reduce some of the ui/ code. > > Anyway, it is probably way to much thrashing, just some ideas to help > promote better data visibilty. > > I was enjoying the ideas of 'groups' and how it can help re-arrange the > data and allow us to look at bottlenecks differently. While --field and > --hierarchy can achieve similar things, I am wondering if the output is > still simple enough to interpret (and the commandline simple enough for > users to utilize). > > My 2cents. Time to jump on a plane. Thanks for your feedback and suggestion. Yes, making output more simple and intuitive is very important. I'll think about how to improve it too. Have a nice flight. Thanks, Namhyung -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/