Thread: Migrated from 8.3 to 9.0 - need to update config (re-post)

Carlo Stonebanks <stonec.register@sympatico.ca> wrote:

> this is a full-time ETL system, with only a handful of actual
> *users* and automated processes over 300 connections running
> *import* programs 24/7

> Intel* Xeon* Processor X5560 (8M Cache, 2.80 GHz, 6.40
> GT/s Intel* QPI) x 2, dual quad core 48 GB RAM
> RAID 10, 6 X 600 GB 15krpm SAS)

So, eight cores and six spindles.  You are probably going to see
*much* better throughput if you route those 300 workers through
about 22 connections.  Use a connection pooler which limits active
transactions to that and queues up requests to start a transaction.

> Sys admin says that battery-backup RAID controller and
> consequent write settings should have no impact on performance.

With only six drives, I your OS, WAL files, indexes, and heap files
are all in the same RAID?  If so, your sys admin is wrong -- you
want the controller configured for write-back (with automatic switch
to write-through on low or failed battery, if possible).

> max_connections = 300

Too high.  Both throughput and latency should improve with correct
use of a connection pooler.

> shared_buffers =
> 500MB        # At 48GB of RAM, could we go to 2GB

You might benefit from as much as 8GB, but only testing with your
actual load will show for sure.

> effective_cache_size =
> 2457MB # Sys admin says assume 25% of 48GB

Add together the shared_buffers setting and whatever the OS tells
you is used for cache under your normal load.  It's usually 75% of
RM or higher.  (NOTE: This doesn't cause any allocation of RAM; it's
a hint to the cost calculations.)

> work_mem =
> 512MB              # Complex reads are called many times a second

Maybe, if you use the connection pooler as described above.  Each
connection can allocate this multiple times.  So with 300
connections you could very easily start using 150GB of RAM in
addition to your shared buffers; causing a swap storm followed by
OOM crashes.  If you stay with 300 connections this *must* be
reduced by at least an order of magnitude.

> # from each connection, so what should this be?
> maintenance_work_mem =
> 256MB  # Should this be bigger - 1GB at least?

I'd go to 1 or 2 GB.

> checkpoint_segments =
> 128     # There is lots of write activity; this is high

OK

> # but could it be higher?

IMO, there's unlikely to be much benefit beyond that.

> #checkpoint_completion_target not set;
> #  Recommendation appears to be .9 for our 128 checkpoint segments

0.9 is probably a good idea.

> default_statistics_target =
> 200  # Deprecated?

Depends on your data.  The default is 100.  You might want to leave
that in general and boost it for specific columns where you find it
is needed.  Higher values improve estimates and can lead to better
query plans, but boost ANALYZE times and query planning time.

> # What is the metric for wal_buffers setting?
> wal_buffers =
> 4MB             # Looks low, recommendation appears to be 16MB.

16MB is good.

> # Is it really "set it and forget it"?

Yeah.

> #synchronous_commit not set;
>
> #  Recommendation is to turn this off and leave fsync on

If this is off, it makes lack of write-back on the controller a lot
less painful.  Even with write-back it can improve performance some.
It does mean that on a crash you can lose some committed
transactions (typically less than a second's worth), but you will
still have database integrity.

> #fsync not set;
>
> #  Recommendation is to leave this on

Unless you want to rebuild your database from scratch or restore
from backup on an OS crash, leave this on.

> #wal_level not set;
>
> #  Do we only needed for replication?

The lowest level just supports crash recovery.  The next level
supports archiving, for recovery from a PITR-style backup.  The
third level is needed to support hot standby (a replicated server on
which you can run targets as it is updated).

> # The issue of vacuum/analyze is a tricky one.
> #  Data imports are running 24/7. One the DB is seeded, the vast
> #  majority of write activity is updates, and not to indexed
> #  columns.  Deletions are vary rare.
> vacuum_cost_delay =
> 20ms

You could try that.  I would monitor for bloat and make things more
aggressive if needed.  If you are not vacuuming aggressively enough,
performance will slowly degrade.  If you let it go too far, recovery
can be a lot of work.

> # The background writer has not been addressed at all.
> #  Can our particular setup benefit from changing the bgwriter
> #  values?

Probably not.  If you find that your interactive users have periods
where queries seem to "freeze" for a few minutes at a time and then
return to normal levels of performance, you might need to make this
more aggressive.

-Kevin

On Fri, Sep 9, 2011 at 5:38 PM, Kevin Grittner
<Kevin.Grittner@wicourts.gov> wrote:
> This is getting back to that issue of using only enough processes at
> one time to keep all the bottleneck resources fully utilized.  Some
> people tend to assuem that if they throw a few more concurrent
> processes into the mix, it'll all get done sooner.  There are a
> great many benchmarks which show otherwise.

On the other hand, in order to benefit from synchro scans and stuff
like that, one has to increase concurrency beyond what is normally
considered optimal.

I have an application that really benefits from synchro scans, but
from time to time, when planets are aligned wrong, the extra
concurrency does hurt.

On Fri, Sep 9, 2011 at 3:16 PM, Kevin Grittner
<Kevin.Grittner@wicourts.gov> wrote:
> Add together the shared_buffers setting and whatever the OS tells
> you is used for cache under your normal load.  It's usually 75% of
> RM or higher.  (NOTE: This doesn't cause any allocation of RAM; it's
> a hint to the cost calculations.)

In the manual[0] it says to take into account the number of concurrent
access to different indices and tables:

"
    Sets the planner's assumption about the effective size of the
disk cache that is available to a single query. This is factored into
estimates of the cost of using an index; a higher value makes it more
likely index scans will be used, a lower value makes it more likely
sequential scans will be used. When setting this parameter you should
consider both PostgreSQL's shared buffers and the portion of the
kernel's disk cache that will be used for PostgreSQL data files. Also,
take into account the expected number of concurrent queries on
different tables, since they will have to share the available space.
This parameter has no effect on the size of shared memory allocated by
PostgreSQL, nor does it reserve kernel disk cache; it is used only for
estimation purposes. The default is 128 megabytes (128MB).
"

However, every mail I've seen on the list, and every bibliography
seems to ignore that. Does PG consider it automatically now, and
admins only have to input the amount of system memory? (in which case
PG could autoconfigure itself by querying /proc), is the manual wrong,
or is the advise given everywher just ignoring that bit?


[0] http://www.postgresql.org/docs/9.0/static/runtime-config-query.html

Claudio Freire <klaussfreire@gmail.com> wrote:
> On Fri, Sep 9, 2011 at 3:16 PM, Kevin Grittner
> <Kevin.Grittner@wicourts.gov> wrote:
>> Add together the shared_buffers setting and whatever the OS tells
>> you is used for cache under your normal load.  It's usually 75%
>> of RM or higher.  (NOTE: This doesn't cause any allocation of
>> RAM; it's a hint to the cost calculations.)
>
> In the manual[0] it says to take into account the number of
> concurrent access to different indices and tables:

Hmm.  I suspect that the manual is technically correct, except that
it probably only matters in terms of how many connections will
concurrently be executing long-running queries which might access
large swaths of large indexes.  In many environments, there are a
lot of maintenance and small query processes, and only occasional
queries where this setting would matter.  I've always had good
results (so far) on the effective assumption that only one such
query will run at a time.  (That is probably helped by the fact that
we normally submit jobs which run such queries to a job queue
manager which runs them one at a time...)

This is getting back to that issue of using only enough processes at
one time to keep all the bottleneck resources fully utilized.  Some
people tend to assuem that if they throw a few more concurrent
processes into the mix, it'll all get done sooner.  There are a
great many benchmarks which show otherwise.

-Kevin

Claudio Freire <klaussfreire@gmail.com> wrote:

> On the other hand, in order to benefit from synchro scans and
> stuff like that, one has to increase concurrency beyond what is
> normally considered optimal.

That's a good example of why any general configuration advice should
just be used as a starting point.  There's no substitute for
benchmarking your own real workload on your own hardware.

On the third hand, though, you have to be very careful about
interpreting these results -- if you used a configuration with a
small effective_cache_size so you could get a lot of benefit from
the synchro scans, you might have suppressed choice of an index
which would have allowed them to run faster with lower concurrency.
Or you might have had to cut your work_mem to a small enough size
(to avoid OOM errors) to force a totally different plan.  So to get
a meaningful comparison, you have to change a number of variables at
once.

Good benchmarking is really hard.

-Kevin

Hi Kevin,

First, thanks for taking the time. I wish I could write back with quick,
terse questions to your detailed reply - but I'm sorry, this is still going
to be a wordy post.

>> max_connections = 300
>Too high.  Both throughput and latency should improve with correct use of
>a connection pooler.

Even for 300 stateful applications that can remain connected for up to a
week, continuously distilling data (imports)? The 300 is overkill, a sys
admin raised it from 100 when multiple large projects were loaded and the
server refused the additional connections. We can take large imports and
break them into multiple smaller ones which the operators are doing to try
and improve import performance. It does result in some improvement, but I
think they have gone over the top and the answer is to improve DB and OS
performance. Perhaps I don't understand how connection pooling will work
with stateful apps that are continuously reading and writing (the apps are
DB I/O bound).

> you want the controller configured for write-back (with automatic switch
> to write-through on low or failed battery, if possible).

For performance or safety reasons? Since the sys admin thinks there's no
performance benefit from this, I would like to be clear on why we should do
this.

>> Can our particular setup benefit from changing the bgwriter values?
> Probably not.  If you find that your interactive users have periods
> where queries seem to "freeze" for a few minutes at a time and then
> return to normal levels of performance, you might need to make this
> more aggressive.

We actually experience this. Once again, remember the overwhelming use of
the system is long-running import threads with continuous connections. Every
now and then the imports behave as if they are suddenly taking a deep
breath, slowing down. Sometimes, so much we cancel the import and restart
(the imports pick up where they left off).

What would the bg_writer settings be in this case?

Thanks again for your time,

Carlo


-----Original Message-----
From: Kevin Grittner [mailto:Kevin.Grittner@wicourts.gov]
Sent: September 9, 2011 2:16 PM
To: pgsql-performance@postgresql.org; Carlo Stonebanks
Subject: Re: [PERFORM] Migrated from 8.3 to 9.0 - need to update config
(re-post)

Carlo Stonebanks <stonec.register@sympatico.ca> wrote:

> this is a full-time ETL system, with only a handful of actual
> *users* and automated processes over 300 connections running
> *import* programs 24/7

> Intel* Xeon* Processor X5560 (8M Cache, 2.80 GHz, 6.40
> GT/s Intel* QPI) x 2, dual quad core 48 GB RAM
> RAID 10, 6 X 600 GB 15krpm SAS)

So, eight cores and six spindles.  You are probably going to see
*much* better throughput if you route those 300 workers through
about 22 connections.  Use a connection pooler which limits active
transactions to that and queues up requests to start a transaction.

> Sys admin says that battery-backup RAID controller and
> consequent write settings should have no impact on performance.

With only six drives, I your OS, WAL files, indexes, and heap files
are all in the same RAID?  If so, your sys admin is wrong -- you
want the controller configured for write-back (with automatic switch
to write-through on low or failed battery, if possible).

> max_connections = 300

Too high.  Both throughput and latency should improve with correct
use of a connection pooler.

> shared_buffers =
> 500MB        # At 48GB of RAM, could we go to 2GB

You might benefit from as much as 8GB, but only testing with your
actual load will show for sure.

> effective_cache_size =
> 2457MB # Sys admin says assume 25% of 48GB

Add together the shared_buffers setting and whatever the OS tells
you is used for cache under your normal load.  It's usually 75% of
RM or higher.  (NOTE: This doesn't cause any allocation of RAM; it's
a hint to the cost calculations.)

> work_mem =
> 512MB              # Complex reads are called many times a second

Maybe, if you use the connection pooler as described above.  Each
connection can allocate this multiple times.  So with 300
connections you could very easily start using 150GB of RAM in
addition to your shared buffers; causing a swap storm followed by
OOM crashes.  If you stay with 300 connections this *must* be
reduced by at least an order of magnitude.

> # from each connection, so what should this be?
> maintenance_work_mem =
> 256MB  # Should this be bigger - 1GB at least?

I'd go to 1 or 2 GB.

> checkpoint_segments =
> 128     # There is lots of write activity; this is high

OK

> # but could it be higher?

IMO, there's unlikely to be much benefit beyond that.

> #checkpoint_completion_target not set;
> #  Recommendation appears to be .9 for our 128 checkpoint segments

0.9 is probably a good idea.

> default_statistics_target =
> 200  # Deprecated?

Depends on your data.  The default is 100.  You might want to leave
that in general and boost it for specific columns where you find it
is needed.  Higher values improve estimates and can lead to better
query plans, but boost ANALYZE times and query planning time.

> # What is the metric for wal_buffers setting?
> wal_buffers =
> 4MB             # Looks low, recommendation appears to be 16MB.

16MB is good.

> # Is it really "set it and forget it"?

Yeah.

> #synchronous_commit not set;
>
> #  Recommendation is to turn this off and leave fsync on

If this is off, it makes lack of write-back on the controller a lot
less painful.  Even with write-back it can improve performance some.
It does mean that on a crash you can lose some committed
transactions (typically less than a second's worth), but you will
still have database integrity.

> #fsync not set;
>
> #  Recommendation is to leave this on

Unless you want to rebuild your database from scratch or restore
from backup on an OS crash, leave this on.

> #wal_level not set;
>
> #  Do we only needed for replication?

The lowest level just supports crash recovery.  The next level
supports archiving, for recovery from a PITR-style backup.  The
third level is needed to support hot standby (a replicated server on
which you can run targets as it is updated).

> # The issue of vacuum/analyze is a tricky one.
> #  Data imports are running 24/7. One the DB is seeded, the vast
> #  majority of write activity is updates, and not to indexed
> #  columns.  Deletions are vary rare.
> vacuum_cost_delay =
> 20ms

You could try that.  I would monitor for bloat and make things more
aggressive if needed.  If you are not vacuuming aggressively enough,
performance will slowly degrade.  If you let it go too far, recovery
can be a lot of work.

> # The background writer has not been addressed at all.
> #  Can our particular setup benefit from changing the bgwriter
> #  values?

Probably not.  If you find that your interactive users have periods
where queries seem to "freeze" for a few minutes at a time and then
return to normal levels of performance, you might need to make this
more aggressive.

-Kevin

On 09/14/2011 02:56 AM, Carlo Stonebanks wrote:

> Even for 300 stateful applications that can remain connected for up to a
> week, continuously distilling data (imports)?

If they're all doing active work all that time you can still benefit
from a pooler.

Say your server can service 50 connections at optimum speed, and any
more result in reduced overall throughput. You have 300 apps with
statements they want to run. Your pooler will basically queue them, so
at any one time 50 are doing work and 250 are waiting for database
access. This should _improve_ database throughput by reducing contention
if 50 worker connections is your sweet spot. However, it will also
increase latency for service for those workers because they may have to
wait a while before their transaction runs, even though their
transaction will complete much faster.

You'd probably want to pool at the transaction level, so once a client
gets a connection it keeps it for the lifetime of that transaction and
the connection is handed back to the pool when the transaction commits
or rolls back.

>>  you want the controller configured for write-back (with automatic
>>  switch to write-through on low or failed battery, if possible).
>
> For performance or safety reasons? Since the sys admin thinks there's no
> performance benefit from this, I would like to be clear on why we should
> do this.

fsync!

If your workload is read-only, it won't help you much. If your workload
is write-heavy or fairly balanced it'll make a HUGE difference, because
fsync() on commit won't have to wait for disk I/O, only I/O to the RAID
card's cache controller.

You can also play with commit_delay and synchronous_commit to trade
guarantees of data persistence off against performance. Don't mind
losing up to 5 mins of commits if you lose power? These options are for you.

Whatever you do, do NOT set fsync=off. It should be called "Eat my data
if anything goes even slightly wrong=on"; it does have legitimate uses,
but they're not yours.

>> > Can our particular setup benefit from changing the bgwriter values?
>>  Probably not. If you find that your interactive users have periods
>>  where queries seem to "freeze" for a few minutes at a time and then
>>  return to normal levels of performance, you might need to make this
>>  more aggressive.
>
> We actually experience this. Once again, remember the overwhelming use
> of the system is long-running import threads with continuous
> connections. Every now and then the imports behave as if they are
> suddenly taking a deep breath, slowing down. Sometimes, so much we
> cancel the import and restart (the imports pick up where they left off).

This could definitely be checkpointing issues. Enable checkpoint logging.

> What would the bg_writer settings be in this case?

You need to tune it for your workload I'm afraid. See the manual and
mailing list discussions.

--
Craig Ringer