Re: System Load analyze

From: "Scott Marlowe" <scott(dot)marlowe(at)gmail(dot)com>
To: "Peter Bauer" <peter(dot)bauer(at)apus(dot)co(dot)at>
Cc: pgsql-general(at)postgresql(dot)org
Subject: Re: System Load analyze
Date: 2007-11-27 17:19:01
Message-ID: dcc563d10711270919o50126e21gf7151e5af2933140@mail.gmail.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-general

On Nov 24, 2007 10:57 AM, Peter Bauer <peter(dot)bauer(at)apus(dot)co(dot)at> wrote:
>
> i have a system here with 2 2.4GHz Xeon Processors, 2GB RAM, ONE Disk on
> a Battery Backed Write Cache SCSI Controller and PostgreSQL 8.1.4
> running with the data on a DRBD Device for High Availability. The used
> database is also replicated to two similar machines with slony1.

Why are you running a version of PostgreSQL with known data eating
bugs? If you care for your data, you will keep up to date on releases.
8.1.10 was released on 2007-09-17. 8.1.4 was released on 2006-05-23.
That's 16 months of bug fixes you're missing. Go here:
http://www.postgresql.org/docs/8.1/static/release.html and read up on
the fixes you're missing. Then update. Or just update.

OK, on the the issue at hand.

> Since the load average is between 1 (most of the time) and 10 (peeks) i
> am worried about the load and executed vmstat and iostat which show that
> 1000-6000 Blocks are writen per second. Please check the attached output
> for further details.
> top shows that the CPUs are at least 80% idle most of the time so i
> think there is an I/O bottleneck. I'm aware that this hardware setup is
> probably not sufficient but is would like to investigate how critical
> the situation is.

Yes. Battery backed cache can only do so much, it's not magic pixie
dust. Once it's full, the drive becomes the bottle neck. Real db
servers have more than one disk drive. They usually have at least 4
or so, and often dozens to hundreds. Also, not all battery backed
caching RAID controllers are created equal.

> procs -----------memory---------- ---swap-- -----io---- --system-- ----cpu----
> r b swpd free buff cache si so bi bo in cs us sy id wa
> 0 0 398256 78328 140612 1063556 0 0 0 1472 2029 5081 4 3 92 0
> 0 2 398256 78268 140612 1063576 0 0 0 2304 1928 4216 0 2 98 0
> 1 2 398256 78100 140612 1063576 0 0 0 1464 1716 3994 1 3 96 0
> 0 0 398256 78704 140612 1063592 0 0 0 916 1435 3570 5 3 91 0
> 0 0 398256 77876 140612 1063616 0 0 0 0 305 1169 3 1 96 0

See that dip in the last line above where the blocks in drop to 0,
idle jumps to 96, and blocks out drops, and context switches drop?
That's most likely where postgresql is checkpointing. Checkpointing
is where it writes out all the dirty buffers to disk. If the bgwriter
is not tuned aggresively enough, checkpoints happen and make the whole
database slow down for a few seconds. If it's tuned too aggresively
then the db spends too much CPU time tracking the dirty buffers and
then writing them. If tuned just right, it will write out the dirty
buffers just fast enough that a checkpoint is never needed.

You tune the bgwriter to your machine and I/O subsystem. If you're
planning on getting more hard drives, do that first. Then tune the
bgwriter.

btw, if this is "vmstat 1" running, it's showing a checkpoint every 20
or so seconds I think

> 0 2 398256 79136 140612 1063964 0 0 0 1736 1959 4494 4 2 94 0
checkpoint here:
> 0 0 398256 79132 140612 1063964 0 0 0 4 260 1039 1 1 98 0
> 0 0 398256 79052 140612 1063980 0 0 0 2444 3084 6955 6 5 89 0
> 0 2 398256 79060 140612 1063988 0 0 0 948 1146 3616 3 1 96 0
> 0 1 398256 78268 140612 1064056 0 0 0 1908 1809 4086 6 5 88 0
> 0 1 398256 76728 140612 1064056 0 0 0 6256 6637 15472 5 5 90 0
> 0 2 398256 77000 140612 1064064 0 0 0 4916 5840 12107 1 4 95 0
> 0 2 398256 76956 140612 1064068 0 0 0 6468 7432 15211 1 3 96 0
> 0 6 398256 77388 140612 1064072 0 0 0 8116 7826 18265 1 8 91 0
> 0 2 398256 74312 140612 1064076 0 0 0 7032 6886 16136 2 7 91 0
> 0 2 398256 74264 140612 1064076 0 0 0 5680 7143 13411 0 5 95 0
> 0 2 398256 72980 140612 1064140 0 0 0 5396 6377 13251 6 6 88 0
> 0 3 398256 76972 140612 1064148 0 0 0 5652 6793 14079 4 9 87 0
> 0 2 398256 77836 140612 1064148 0 0 0 3968 5321 14187 10 8 82 0
> 1 0 398256 77280 140612 1064148 0 0 0 1608 3188 8974 21 12 67 0
> 1 0 398256 77832 140612 1064152 0 0 0 236 834 2625 7 5 87 0
> 0 0 398256 77464 140612 1064152 0 0 0 244 505 1378 2 4 94 0
> 1 0 398256 77828 140612 1064164 0 0 0 316 580 1954 7 2 91 0
> 0 0 398256 77804 140612 1064180 0 0 0 740 673 2248 2 2 96 0
> 0 0 398256 77000 140612 1064180 0 0 0 304 589 1739 1 3 96 0
20 rows later, checkpoint here:
> 0 0 398256 77000 140612 1064184 0 0 0 0 216 886 0 1 99 0
> 0 0 398256 75452 140612 1064184 0 0 0 432 755 2032 6 1 93 0

> max_fsm_pages = 40000 # min max_fsm_relations*16, 6 bytes each, APUS

This seems a little low for a busy server.

> # - Background writer -
>
> bgwriter_delay = 100 # 10-10000 milliseconds between rounds, APUS
> bgwriter_lru_percent = 2.0 # 0-100% of LRU buffers scanned/round, APUS
> bgwriter_lru_maxpages = 10 # 0-1000 buffers max written/round, APUS
> bgwriter_all_percent = 1 # 0-100% of all buffers scanned/round, APUS
> bgwriter_all_maxpages = 10 # 0-1000 buffers max written/round, APUS

So, bgwriter wakes up 10 times a second, and each time it processes 2%
of the Least Recently Used pages for writing, and writes up to 10 of
those pages. And it only checks 1% of the total pages and writes 10
of those at the most. This is not aggresive enough, and given how
much spare CPU you have left over, you can be a fair bit more
aggresive. The main thing to increase is the maxes. Try changing
them to the 100 to 300 range, and maybe increase your percentages to
5% or so. What we're shooting for is to see those checkpoints go
away.

Then, when running your benchmark, after a few minutes, run a
checkpoint by hand and see if you get one of those slow downs like we
saw in vmstat above. If your bgwriter is tuned properly, you should
get an almost instant response from the checkpoint and no noticeable
slow down in the vmstat numbers for context switches per second.

Once you reach the point where the bgwriter is just keeping ahead of
check points, there's little to be gained in more aggressive tuning of
the bgwriter and you'll just be chewing up memory and cpu bandwidth if
you do get too aggressive with it.

In response to

Responses

Browse pgsql-general by date

  From Date Subject
Next Message Tom Lane 2007-11-27 17:19:51 Re: [GENERAL] Empty arrays with ARRAY[]
Previous Message Filip Rembiałkowski 2007-11-27 17:18:18 Re: postgres 8.3 beta 2 storage question