Details about pg_stat_bgwriter

[Thomas Kellerer]

Hi,

I'm trying to understand the information shown in pg_stat_bgwriter.

If I want to identify a possible bottleneck with the bgwriter, for what kind of numbers do I need to watch out?

Are the following values OK, or do they indicate that I need to tune my configuration?

checkpoints_timed : 6099
checkpoints_req : 291
buffers_checkpoint : 2767641
buffers_clean : 4522
maxwritten_clean : 1
buffers_backend : 609933
buffers_alloc : 1411947

I have read Greg's description (http://www.westnet.com/~gsmith/content/postgresql/chkp-bgw-83.htm), but to be honest I don't understand the explanation. He mentions a "LRU background writer", but I'm not sure to which column that relates.

Thanks in advance.
Thomas

[Greg Smith]

Post by Thomas Kellerer
If I want to identify a possible bottleneck with the bgwriter, for
what kind of numbers do I need to watch out?

You don't much with a single snapshot of pg_stat_bgwriter data. Try
saving this instead:

select *,now() from pg_stat_bgwriter;

And then take another snapshot at least a few hours later, preferably
the next day. With two snapshots and timestamps on them, then it's
possible to make some sense of the numbers.
--
Greg Smith 2ndQuadrant US Baltimore, MD
PostgreSQL Training, Services and Support
***@2ndQuadrant.com www.2ndQuadrant.us

[Thomas Kellerer]

Post by Thomas Kellerer
If I want to identify a possible bottleneck with the bgwriter, for
what kind of numbers do I need to watch out?

You don't much with a single snapshot of pg_stat_bgwriter data. Try
select *,now() from pg_stat_bgwriter;
And then take another snapshot at least a few hours later, preferably
the next day. With two snapshots and timestamps on them, then it's
possible to make some sense of the numbers.

Thanks for the answer.

After taking a few snapshots, what would I look for?

Regards
Thomas

[Greg Smith]

Post by Greg Smith
You don't much with a single snapshot of pg_stat_bgwriter data. Try
select *,now() from pg_stat_bgwriter;
And then take another snapshot at least a few hours later, preferably
the next day. With two snapshots and timestamps on them, then it's
possible to make some sense of the numbers.

I probably should have explained the next part. I've now shared what I
do with this information at
http://www.pgcon.org/2010/schedule/events/218.en.html

Basically, if you put the data from the two snapshots into one of the
Statistics Spreadsheet versions, you'll get several derived numbers that
pop out:

-Average checkpoint frequency
-Average size of each checkpoint
-Average rate at which new buffers are allocated
-Average rate of writes out of the buffer cache
-Percentage of writes done by checkpoints, the background writer LRU
cleaner, and client backends

These are the sort of things you can actually think about in useful
real-world terms. And if you tune the database by doing things like
increasing checkpoint_segments/checkpoint_timeout or changing the size
of shared_buffers, you can see what impact that has on the overall
system response, by noting how it changes these numeric measures. It's
a bit more exact of an approach for tuning such things than what people
have traditionally done here. Increasing shared_buffers and watching
the total writes/second number drop significantly is more exact than
just nothing whether the server waiting for I/O percentage dropped or not.

I also highly recommend installing pg_buffercache and looking at what's
inside your buffer cache too--the data actually being cached by the
server never fails to surprise people, relative to what they expected
was being cached. Source for how to analyze those things is in the
pg_buffercache samples shell script, the "Buffer contents summary, with
percentages" is the one that gives useful output. One of the thing I
did in the presentation that does not come through on the slides is note
how the distribution of high usage count data in your buffer cache
determines whether you will benefit from making it large or not. I've
reduced this to a single figure of merit now: "cumulative % of buffers
that have a usage count of >=2". On a system that benefits from having
a big buffer cache, that number will be quite large (86% in the first of
the samples in the spreadsheet). On one that doesn't, it will be small
(35% on the last system listed there). You really have to tune those
two types of workload differently.
--
Greg Smith 2ndQuadrant US Baltimore, MD
PostgreSQL Training, Services and Support
***@2ndQuadrant.com www.2ndQuadrant.us

[Scott Marlowe]

You don't much with a single snapshot of pg_stat_bgwriter data.  Try
select *,now() from pg_stat_bgwriter;
And then take another snapshot at least a few hours later, preferably the
next day.  With two snapshots and timestamps on them, then it's possible to
make some sense of the numbers.

I probably should have explained the next part.  I've now shared what I do
with this information at
http://www.pgcon.org/2010/schedule/events/218.en.html
Basically, if you put the data from the two snapshots into one of the
Statistics Spreadsheet versions, you'll get several derived numbers that pop
-Average checkpoint frequency
-Average size of each checkpoint
-Average rate at which new buffers are allocated
-Average rate of writes out of the buffer cache
-Percentage of writes done by checkpoints, the background writer LRU
cleaner, and client backends

I think you get all or most of that if you just log checkpoints.

[Greg Smith]

Post by Scott Marlowe

Post by Greg Smith
-Average checkpoint frequency
-Average size of each checkpoint
-Average rate at which new buffers are allocated
-Average rate of writes out of the buffer cache
-Percentage of writes done by checkpoints, the background writer LRU
cleaner, and client backends

I think you get all or most of that if you just log checkpoints.

If you turned on log_checkpoints, and you went through a whole stack of
logs summarizing the data it writes out, you can derive the first two of
those from it. I happen to think that sampling two data points and
pasting into a spreadsheet is less hassle to deal with, and the data is
definitely there by default. You can answer questions like "what's the
biggest checkpoint I've seen?" from the stuff in the logs a bit easier
than from pg_stat_bgwriter.

After for the rest, checkpoints are one of the three possible ways that
buffers can be written out. The other two are the background writer's
cleaning and client backends. log_checkpoints gives you the size of one
of those sources of writes. You can also compute an estimated
proportion between the three types of writes from a single snapshot of
pg_stat_bgwriter data. But that doesn't help once you've reached the
point where you want to change something and measure how response
changes afterwards. That requires more regularly sampling the data, so
you have a delta between two times.
--
Greg Smith 2ndQuadrant US Baltimore, MD
PostgreSQL Training, Services and Support
***@2ndQuadrant.com www.2ndQuadrant.us

[Thomas Kellerer]

I probably should have explained the next part. I've now shared what I
do with this information at
http://www.pgcon.org/2010/schedule/events/218.en.html
Basically, if you put the data from the two snapshots into one of the
Statistics Spreadsheet versions, you'll get several derived numbers that
-Average checkpoint frequency
-Average size of each checkpoint
-Average rate at which new buffers are allocated
-Average rate of writes out of the buffer cache
-Percentage of writes done by checkpoints, the background writer LRU
cleaner, and client backends
These are the sort of things you can actually think about in useful
real-world terms. And if you tune the database by doing things like
increasing checkpoint_segments/checkpoint_timeout or changing the size
of shared_buffers, you can see what impact that has on the overall
system response, by noting how it changes these numeric measures. It's a
bit more exact of an approach for tuning such things than what people
have traditionally done here. Increasing shared_buffers and watching the
total writes/second number drop significantly is more exact than just
nothing whether the server waiting for I/O percentage dropped or not.
I also highly recommend installing pg_buffercache and looking at what's
inside your buffer cache too--the data actually being cached by the
server never fails to surprise people, relative to what they expected
was being cached. Source for how to analyze those things is in the
pg_buffercache samples shell script, the "Buffer contents summary, with
percentages" is the one that gives useful output. One of the thing I did
in the presentation that does not come through on the slides is note how
the distribution of high usage count data in your buffer cache
determines whether you will benefit from making it large or not. I've
reduced this to a single figure of merit now: "cumulative % of buffers
that have a usage count of >=2". On a system that benefits from having a
big buffer cache, that number will be quite large (86% in the first of
the samples in the spreadsheet). On one that doesn't, it will be small
(35% on the last system listed there). You really have to tune those two
types of workload differently.

Thanks a lot for this detailed information!

I don't think we have a caching issue though, it's just one (regular) bulk import that does some heavy writes which we need to tune. Read performance is absolutely fine, but everything around vacuum and checkpointing seems to slow down the import massively.

So would taking a snapshot before and after this import be the right way to start the analysis?

Regards
Thomas

[Greg Smith]

Post by Thomas Kellerer
I don't think we have a caching issue though, it's just one (regular)
bulk import that does some heavy writes which we need to tune. Read
performance is absolutely fine, but everything around vacuum and
checkpointing seems to slow down the import massively.
So would taking a snapshot before and after this import be the right
way to start the analysis?

Sure. Might want to grab one at those two points, then at a point right
before the next big import happens, so that you have what a relatively
idle period looks like for comparison sake.
--
Greg Smith 2ndQuadrant US Baltimore, MD
PostgreSQL Training, Services and Support
***@2ndQuadrant.com www.2ndQuadrant.us