Re: BDR, wal sender, high system cpu, mutex_lock_common

On 12 October 2017 at 11:03, Andres Freund <andres(at)anarazel(dot)de> wrote:
> On 2017-10-12 10:25:43 +0800, Craig Ringer wrote:
>> On 4 October 2017 at 00:21, milist ujang <ujang(dot)milist(at)gmail(dot)com> wrote:
>> > On Tue, Oct 3, 2017 at 8:49 PM, Craig Ringer <craig(at)2ndquadrant(dot)com> wrote:
>> >>
>> >>
>> >> Can you get stacks please?
>> >>
>> >> Use -g
>> >
>> >
>> > # Events: 2K cpu-clock
>> > #
>> > # Overhead Command Shared Object Symbol
>> > # ........ ........ ................. ............................
>> > #
>> > 86.96% postgres [kernel.kallsyms] [k] __mutex_lock_common.isra.5
>> > |
>> > --- __mutex_lock_common.isra.5
>> > read
>>
>>
>> Unfortunately it looks like you're using a postgres built with
>> -fomit-frame-pointers (the default) on x64, with an older perf not
>> built with libunwind. This produces useless stacks.
>
> Just read this mail, but for libunwind to work you'd have to specify
> "--call-graph dwarf", no?

I think you're right. But only on a version of perf where it's
available and used.

I haven't recently checked if perf has finally grown the ability to
load external debug symbols either. It never used to.

--
Craig Ringer http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services