RE: LWLocks by LockManager slowing large DB

From: Paul Friedman <paul(dot)friedman(at)streetlightdata(dot)com>
To: Andres Freund <andres(at)anarazel(dot)de>
Cc: pgsql-performance(at)lists(dot)postgresql(dot)org
Subject: RE: LWLocks by LockManager slowing large DB
Date: 2021-04-12 22:56:08
Message-ID: cff944a561ad1e8f5d08a5df5f8f0a99@mail.gmail.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-performance

Yes, I ran the query after a couple of minutes. Those are the
steady-state numbers.

Also 'top' shows:

top - 22:44:26 up 12 days, 23:14, 5 users, load average: 20.99, 21.35,
19.27
Tasks: 859 total, 26 running, 539 sleeping, 0 stopped, 0 zombie
%Cpu(s): 34.3 us, 1.6 sy, 0.0 ni, 64.1 id, 0.0 wa, 0.0 hi, 0.0 si,
0.0 st
KiB Mem : 24742353+total, 33723356 free, 73160656 used,
14053952+buff/cache
KiB Swap: 0 total, 0 free, 0 used. 17132937+avail Mem

PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+
COMMAND

30070 postgres 20 0 41.232g 28608 18192 R 53.8 0.0 17:35.74
postgres

30087 postgres 20 0 41.233g 28408 18180 S 53.8 0.0 17:35.69
postgres

30055 postgres 20 0 41.233g 28492 18120 R 53.5 0.0 17:41.51
postgres

Note the postgres processes only running at 53% with the system at 64%
idle. The 1.7% system time seems indicative of the spinlocks blocking the
major processing.

Do you know what resource the LockManager might be blocking on/protecting
with these LWlocks?

Also, I didn't understand your comment about a 'futex profile', could you
point me in the right direction here?

Thanks.

---Paul

Paul Friedman
CTO

677 Harrison St  |  San Francisco, CA 94107
M: (650) 270-7676
E-mail: paul(dot)friedman(at)streetlightdata(dot)com

-----Original Message-----
From: Andres Freund <andres(at)anarazel(dot)de>
Sent: Monday, April 12, 2021 3:22 PM
To: Paul Friedman <paul(dot)friedman(at)streetlightdata(dot)com>
Cc: pgsql-performance(at)lists(dot)postgresql(dot)org
Subject: Re: LWLocks by LockManager slowing large DB

Hi,

On 2021-04-12 15:15:05 -0700, Paul Friedman wrote:
> Thanks again for any advice you have.

I think we'd need the perf profiles to be able to dig into this further.

It's odd that there are a meaningful amount of LockManager contention in
your case - assuming the stats you collected weren't just the first few
milliseconds of starting those 60 queries, there shouldn't be any
additional "heavyweight locks" taken given the duration of your queries.

The futex profile hopefully will tell us from where that is coming from...

Greetings,

Andres Freund

In response to

Responses

Browse pgsql-performance by date

  From Date Subject
Next Message Andres Freund 2021-04-12 23:03:54 Re: LWLocks by LockManager slowing large DB
Previous Message Andres Freund 2021-04-12 22:22:21 Re: LWLocks by LockManager slowing large DB