vmdaemon CPU usage and poor performance in 10.0-RELEASE

[Alan Cox]

On Fri, Aug 15, 2014 at 3:43 AM, Ronald Klop <ronald-lists at klop.ws> wrote:

> On Wed, 13 Aug 2014 17:42:37 +0200, Polyack, Steve <
> Steve.Polyack at intermedix.com> wrote:
>
>
>>  -----Original Message-----
>>> From: Kurt Jaeger [mailto:lists at opsec.eu]
>>> Sent: Wednesday, August 13, 2014 11:25 AM
>>>
>>> Hi!
>>>
>>> > We have a handful of database servers running FreeBSD 10.0-RELEASE
>>> > and PostgreSQL 9.3.4.  The servers have 128GB or 256GB of RAM.
>>>
>>> Are you aware of the recent work on that topic ?
>>>
>>> https://www.freebsd.org/news/status/report-2014-04-2014-
>>> 06.html#PostgreSQL-Performance-Improvements
>>>
>>> Maybe kib@ knows more about this ?
>>>
>>>
>> I've recently read over this and some other posts, but they all seem to
>> center around poor postgres performance.  In our case at least, some light
>> to medium usage of postgres generally makes the entire system unusable.
>>
>> The patches & documents linked there also all seem to be for -CURRENT,
>> which we aren't running.  We're not too keen on the idea of using CURRENT
>> in production, either.  We're planning on testing 10-STABLE, but I was just
>> hoping to gain some insight into what the problem may be and whether recent
>> commits to vmdaemon code in the -STABLE tree may have a positive effect on
>> what we've seen.
>>
>> Steve
>>
>
> It looks like a fix mentioned in part 2.1 in the pdf (
> https://www.kib.kiev.ua/kib/pgsql_perf_v2.0.pdf, from the status report)
> was only just committed to 11-CURRENT.
> http://svnweb.freebsd.org/changeset/base/270011
>
> I guess it is advisable to stay on pgsql 9.2.x until these improvements
> are MFC'ed to 10-STABLE.
>
>

The issues discussed in Kostik's report have nothing to do with the page
daemon or the problem that started this thread.  10.0 had a regression in
page daemon operation that I fixed in r265945.

Alan