Обсуждение: time to stop tuning?

We have a network application in which many clients will be executing a mix of select/insert/update/deletes on a
centralpostgres 7.4.5 database, running on Solaris 9 running on dual 2.3 ghz Xeons, with 2 gig of RAM and a RAID 10
disk.The test database is about 400 meg in size. 

We have tuned the postgresql.conf parameters to the point where we are confident we have enough memory for shared
buffersand for sorting. We are still tuning SQL statements, but we're pretty sure the big wins have been achieved. 

We are maxing out on the backend with 30 postmaster processes, each taking up about 2.5-3% of the CPU. We have tested
mountingthe whole database in /tmp, hence in memory, and it has made no difference in performance, so it seems we are
purelyCPU bound at this point. 

About 70% of our time is spent in selects, and another 25% spent in inserts/updates of a single table (about 10% out of
theselects % is against this table). 

Now, our application client is not doing nearly enough of it's own caching, so a lot the work the database is doing
currentlyis redundant, and we are working on the client, but in the meantime we have to squeeze as much as we can from
thebackend. 

After that long intro, I have a couple of questions:

1) Given that the data is all cached, what can we do to make sure that postgres is generating
the most efficient plans in this case? We have bumped up effective_cache_size, but it had no
effect. Also, what would the most efficient plan for in-memory data look like? I mean, does one
still look for the normal stuff - index usage, etc., or are seqscans what we should be looking for?
I've seen some stuff about updating statistics targets for specific tables, but I'm not sure I
understand it, and don't know if something like that applies in this case. I can supply some specific plans, if that
wouldhelp (this email is already too long...). 

2) We have SQL test environment where we just run the SQL statements executed by the clients (culled from the log file)
inpsql. In our test environment, the same set of SQL statements runs 4X faster that the times achieved in the test that
generatedour source log file. Obviously there was a bigger load on the machine in the full test, but I'm wondering if
thereare any particular diagnostics that I should be looking at to ferret out contention. I haven't seen anything that
lookedsuspicious in pg_locks, but it's difficult to interpret that data when the database is under load (at least for
someoneof my limited experience). 

I suspect the ultimate answer to our problem will be:

   1) aggressive client-side caching
   2) SQL tuning
   3) more backend hardware

But I would grateful to hear any tips/anecdotes/experiences that others might have from tuning similar applications.

Thanks!

- DAP
----------------------------------------------------------------------------------
David Parker    Tazz Networks    (401) 709-5130
 

"David Parker" <dparker@tazznetworks.com> writes:
> 1) Given that the data is all cached, what can we do to make sure that
> postgres is generating the most efficient plans in this case? We have
> bumped up effective_cache_size, but it had no effect.

If you're willing to bet on everything being in RAM all the time,
dropping random_page_cost to 1 would be a theoretically sound thing
to do.  In any case you should look at reducing it considerably from
the default setting of 4.

Something that might also be interesting is to try increasing all the
cpu_xxx cost factors, on the theory that since the unit of measurement
(1 sequential page fetch) relates to an action involving no actual I/O,
the relative costs of invoking an operator, etc, should be rated higher
than when you expect actual I/O.  I'm not real confident that this would
make things better --- you might find that any improvement would be
swamped by the low accuracy with which we model CPU costs (such as the
assumption that every operator costs the same to evaluate).  But it's
worth some experimentation.

            regards, tom lane

Hi. Thanks for responding.

As it happens, the client-side already has a connection pool.

We need statistics enabled so that autovacuum can run (without
autovacuum running our updates begin to kill us pretty quickly).

Moving off of Solaris 9 isn't an option, even for the purposes of
comparison, unfortunately.

On limiting the client side connections: we've been gradually pushing up
the client-side connection pool and threads, and have seen steady
improvement in our throughput up to the current barrier we have reached.
I guess the idea would be that backing off on the connections would
allow each operation to finish faster, but that hasn't been the observed
behavior so far.

I've attached the plans for the 4 queries that represent ~35% of our
load. These are run against the same dataset, but without any other
load. Another big query basically requires a test to be runnning because
the data is transient, and I can't run that at the moment. The times for
the individual queries is really fine - it's just they are called 3
times for every logical "unit of work" on the client side, so they are
called thousands of times in a given test (hence the need for client
caching).

Thanks.

- DAP

>-----Original Message-----
>From: Rod Taylor [mailto:rbt@sitesell.com]
>Sent: Friday, November 26, 2004 1:29 PM
>To: David Parker
>Cc: Postgresql Performance
>Subject: Re: [PERFORM] time to stop tuning?
>
>On Fri, 2004-11-26 at 12:13 -0500, David Parker wrote:
>>
>> I suspect the ultimate answer to our problem will be:
>>
>>    1) aggressive client-side caching
>>    2) SQL tuning
>>    3) more backend hardware
>
>#0 might actually be using connection pooling and using cached
>query plans (PREPARE), disabling the statistics daemon, etc.
>
>For the plans, send us EXPLAIN ANALYZE output for each of the
>common queries.
>
>If you can try it, I'd give a try at FreeBSD or a newer Linux
>on your system instead of Solaris. Older versions of Solaris
>had not received the same amount of attention for Intel
>hardware as the BSDs and Linux have and I would imagine
>(having not tested it recently) that this is still true for
>32bit Intel.
>
>Another interesting test might be to limit the number of
>simultaneous connections to 8 instead of 30 (client side
>connection retry) after client side connection pooling via
>pgpool or similar has been installed.
>
>Please report back with your findings.
>--
>Rod Taylor <rbt@sitesell.com>
>
>

On Fri, 2004-11-26 at 12:13 -0500, David Parker wrote:
>
> I suspect the ultimate answer to our problem will be:
>
>    1) aggressive client-side caching
>    2) SQL tuning
>    3) more backend hardware

#0 might actually be using connection pooling and using cached query
plans (PREPARE), disabling the statistics daemon, etc.

For the plans, send us EXPLAIN ANALYZE output for each of the common
queries.

If you can try it, I'd give a try at FreeBSD or a newer Linux on your
system instead of Solaris. Older versions of Solaris had not received
the same amount of attention for Intel hardware as the BSDs and Linux
have and I would imagine (having not tested it recently) that this is
still true for 32bit Intel.

Another interesting test might be to limit the number of simultaneous
connections to 8 instead of 30 (client side connection retry) after
client side connection pooling via pgpool or similar has been installed.

Please report back with your findings.
--
Rod Taylor <rbt@sitesell.com>

> On limiting the client side connections: we've been gradually pushing up
> the client-side connection pool and threads, and have seen steady
> improvement in our throughput up to the current barrier we have reached.

Very well.. Sometimes more simultaneous workers helps, other times it
hinders.

> I've attached the plans for the 4 queries that represent ~35% of our
> load. These are run against the same dataset, but without any other
> load. Another big query basically requires a test to be runnning because

Those aren't likely from your production system as there isn't any data
in those tables and the queries took less than 1ms.

--
Rod Taylor <rbt@sitesell.com>