Re: [PATCH 1/1] perf tools: perf diff for different binaries
From: Namhyung Kim
Date: Tue Nov 04 2014 - 00:41:48 EST
Hi kan,
On Fri, 31 Oct 2014 05:06:12 -0700, kan liang wrote:
> From: Kan Liang <kan.liang@xxxxxxxxx>
>
> perf diff can display the differential profile between two perf.data
> files. However, the perf.data files have to come from same binaries.
>
> The patch introduced a key "symbol_name" for --sort option, so the user
> can compare the perf.data files from different binaries, even different
> kernel version. If symbol_name is set, perf diff display the
> differential profile of functions.
> It could help the user quickly locate the scaling regression caused by
> new binaries/kernels.
>
> Here is an example. I have two version of binaries. One is running on
> kernel 3.16. The other is running on kernel 3.17.
>
> Using perf record, I got two perf.data files, v1_1_6perf.data (version
> one in 3.16) and v2_1_7perf.data (version two in 3.17).
>
> If I run old perf diff with --sort dso,symbol, we cannot get any useful
> information for the user space binary. Furthermore, it failed to compare
> the kernel function. E.g. native_write_msr_safe.
>
> perf diff -s dso,symbol --compute ratio v1_1_6perf.data v2_1_7perf.data
>
> Event 'cycles'
>
> Baseline Ratio Shared Object Symbol
> ........ .............. .................
> ....................................
>
> [kernel.kallsyms] [k]
> __update_entity_load_avg_contrib
> [kernel.kallsyms] [k] _raw_spin_lock
> [kernel.kallsyms] [k] apic_timer_interrupt
> [kernel.kallsyms] [k] hrtimer_interrupt
> 0.01% [kernel.kallsyms] [k] native_write_msr_safe
> [kernel.kallsyms] [k] native_write_msr_safe
> 0.01% [kernel.kallsyms] [k] notifier_call_chain
> 0.01% [kernel.kallsyms] [k] perf_event_task_tick
> 0.01% [kernel.kallsyms] [k] run_posix_cpu_timers
> 0.01% [kernel.kallsyms] [k] run_timer_softirq
> 0.01% [kernel.kallsyms] [k] trigger_load_balance
> 0.01% [kernel.kallsyms] [k] update_vsyscall
> 0.05% [unknown] [.] 0x0000000000400540
> 0.04% [unknown] [.] 0x0000000000400541
> 0.03% [unknown] [.] 0x000000000040054b
> 0.04% [unknown] [.] 0x0000000000400552
> 33.55% [unknown] [.] 0x0000000000400554
> 1.22% [unknown] [.] 0x000000000040055a
> 8.00% [unknown] [.] 0x000000000040055e
> 0.02% [unknown] [.] 0x0000000000400562
> 8.41% [unknown] [.] 0x0000000000400564
> 48.13% [unknown] [.] 0x000000000040056b
> 0.16% [unknown] [.] 0x0000000000400570
> 0.17% [unknown] [.] 0x0000000000400571
> [unknown] [.] 0x0000000000400580
> [unknown] [.] 0x0000000000400581
> 0.01% [unknown] [.] 0x0000000000400583
> 0.01% [unknown] [.] 0x0000000000400588
> [unknown] [.] 0x000000000040058b
> 0.01% 1240.990221 [unknown] [.] 0x000000000040058d
> [unknown] [.] 0x0000000000400590
> 0.06% [unknown] [.] 0x0000000000400591
> [unknown] [.] 0x0000000000400593
> 0.04% [unknown] [.] 0x0000000000400595
> 0.01% 1240.603148 [unknown] [.] 0x0000000000400597
> [unknown] [.] 0x000000000040059b
> [unknown] [.] 0x000000000040059d
> [unknown] [.] 0x00000000004005a1
> [unknown] [.] 0x00000000004005a5
> [unknown] [.] 0x00000000004005a7
> [unknown] [.] 0x00000000004005a8
> [unknown] [.] 0x00000000004005aa
> [unknown] [.] 0x00000000004005ba
> [unknown] [.] 0x00000000004005bf
> [unknown] [.] 0x00000000004005c4
> [unknown] [.] 0x00000000004005c9
> [unknown] [.] 0x00000000004005ce
> [unknown] [.] 0x00000000004005d2
> [unknown] [.] 0x00000000004005d6
> [unknown] [.] 0x00000000004005d8
> [unknown] [.] 0x00000000004005f5
>
> With the key "symbol_name", we can easily get the differential profile
> from different binary versions and different kernel versions.
>
> ./perf diff -s dso,symbol_name --compute ratio v1_1_6perf.data v2_1_7perf.data
>
> Event 'cycles'
>
> Baseline Ratio Shared Object Symbol Name
> ........ .............. .................
> ....................................
>
> [kernel.kallsyms] [k]
> __update_entity_load_avg_contrib
> [kernel.kallsyms] [k] _raw_spin_lock
> [kernel.kallsyms] [k] apic_timer_interrupt
> [kernel.kallsyms] [k] hrtimer_interrupt
> 0.01% 1.998522 [kernel.kallsyms] [k] native_write_msr_safe
> 0.01% [kernel.kallsyms] [k] notifier_call_chain
> 0.01% [kernel.kallsyms] [k] perf_event_task_tick
> 0.01% [kernel.kallsyms] [k] run_posix_cpu_timers
> 0.01% [kernel.kallsyms] [k] run_timer_softirq
> 0.01% [kernel.kallsyms] [k] trigger_load_balance
> 0.01% [kernel.kallsyms] [k] update_vsyscall
> tchain_edit [.] f1
> 0.14% 3.913444 tchain_edit [.] f2
> 99.82% 1.005478 tchain_edit [.] f3
Hmm.. I think it should be a default behavior for perf diff, otherwise
-s symbol is almost meaningless IMHO. What about setting the
sort_sym.se_collapse in data_process() so that hists__match() can use
symbol names?
Thanks,
Namhyung
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/