On 2020/08/20 20:01, Fujii Masao wrote:
>
>
> On 2020/08/19 14:10, Masahiro Ikeda wrote:
>> On 2020-08-19 13:49, tsunakawa.takay@fujitsu.com wrote:
>>> From: Masahiro Ikeda <ikedamsh@oss.nttdata.com>
>>>> If my understanding is correct, we have to measure the performance
>>>> impact first.
>>>> Do you know HariBabu is now trying to solve it? If not, I will try to
>>>> modify patches to apply HEAD.
>>>
>>> No, he's not doing it anymore. It'd be great if you could resume it.
>>
>> OK, thanks.
>>
>>> However, I recommend sharing your understanding about what were the
>>> issues with those two threads and how you're trying to solve them.
>>> Was the performance overhead the blocker in both of the threads?
>>
>> In my understanding, some comments are not solved in both of the threads.
>> I think the following works are remained.
>>
>> 1) Modify patches to apply HEAD
>> 2) Get consensus what metrics we collect and how to use them for tuning.
>
> I agree to expose the number of WAL write caused by full of WAL buffers.
> It's helpful when tuning wal_buffers size. Haribabu separated that number
> into two fields in his patch; one is the number of WAL write by backend,
> and another is by background processes and workers. But I'm not sure
> how useful such separation is. I'm ok with just one field for that number.
Just idea; it may be worth exposing the number of when new WAL file is
created and zero-filled. This initialization may have impact on
the performance of write-heavy workload generating lots of WAL. If this
number is reported high, to reduce the number of this initialization,
we can tune WAL-related parameters so that more "recycled" WAL files
can be hold.
Regards,
--
Fujii Masao
Advanced Computing Technology Center
Research and Development Headquarters
NTT DATA CORPORATION
