I redid my previous measurements after finishing up the weekend's
hacking. The numbers shown below are elapsed time in seconds for

time psql -f testfile.sql postgres >/dev/null

using CVS HEAD and REL8_1_STABLE branch tip, compiled --enable-debug
--disable-cassert, no nondefault options except for turning fsync off
(which doesn't particularly affect read-only tests like these anyway).
The machines are both running current Fedora Core 5. The older x86
machine is known to have the slow-gettimeofday problem from previous
experimentation with EXPLAIN ANALYZE. Each number is the median of 3
or more tests, rounded off to 0.1 second (I wouldn't put a lot of faith
in differences of 0.1 sec or so, because of the variance I saw in the
tests).

x86 x86_64

8.1 HEAD 8.1 HEAD

100000 "SELECT 1;" 25.9 27.0 9.2 9.1
with stats_command_string=1 63.5 27.6 18.7 9.2
with log_min_duration_statement=100 26.9 27.8 9.6 9.2
with statement_timeout=100 27.5 28.6 9.6 9.8
with all 3 features 66.1 29.3 19.5 9.7

BEGIN, 100000 "SELECT 1;", COMMIT 21.2 23.1 8.3 8.4
with stats_command_string=1 52.3 23.5 15.4 8.5
with log_min_duration_statement=100 22.1 23.4 8.4 8.4
with statement_timeout=100 23.7 24.3 8.6 8.8
with all 3 features 55.2 25.5 16.0 8.8

I chose the log_min_duration_statement and statement_timeout settings
high enough so that no actual logging or timeout would happen --- the
point is to measure the measurement overhead.

The good news is that we've pretty much licked the problem of
stats_command_string costing an unreasonable amount.

The bad news is that except in the stats_command_string cases, HEAD
is noticeably slower than 8.1 on the machine with slow gettimeofday.
In the single-transaction test this might be blamed on the addition
of statement_timestamp support (which requires a gettimeofday per
statement that wasn't there in 8.1) ... but in the one-transaction-
per-statement tests that doesn't hold water, because each branch is
doing a gettimeofday per statement, just in different places.

Can anyone else reproduce this slowdown? It might be only an artifact
of these particular builds, but it's a bit too consistent in my x86 data
to just ignore.

BTW, according to "top" the CPU usage percentages in these tests are
on the order of 55% backend, 45% psql. Methinks psql needs a round
of performance tuning ...

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 6: explain analyze is your friend

I wrote:
> BTW, according to "top" the CPU usage percentages in these tests are
> on the order of 55% backend, 45% psql. Methinks psql needs a round
> of performance tuning ...

oprofile tells the tale:

CPU: P4 / Xeon with 2 hyper-threads, speed 2793.03 MHz (estimated)
Counted GLOBAL_POWER_EVENTS events (time during which processor is not stopped)
with a unit mask of 0x01 (mandatory) count 240000
GLOBAL_POWER_E...|
samples| %|
------------------
682534 52.7683 /usr/lib/debug/lib/modules/2.6.16-1.2133_FC5/vmlinux
274747 21.2413 /home/tgl/testversion/bin/postgres
226306 17.4962 /lib64/libc-2.4.so
54296 4.1977 /home/tgl/testversion/bin/psql
45376 3.5081 /home/tgl/testversion/lib/libpq.so.5.0
5302 0.4099 /usr/bin/oprofiled
1954 0.1511 /oprofile

It's all about the kernel process-switch overhead, which is being blamed
equally on both processes.

I did find some low-hanging fruit in GetVariable(), but nothing else
that looked readily improvable.

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 4: Have you searched our list archives?

http://archives.postgresql.org

Sorry I am traveling for EnterpriseDB Wednesday and Thursday so I can't
run the tests right now.

Seeing stats_command_string with almost zero overhead is great news!
Should we remove that setting and just have it enabled all
the time?

As far as pg_stat_activity.query_start, I never suspected that column
would show the time of start of IDLE. I think we can just document that
that column is the time of the start of the last query, and avoid the
gettimeofday() call on IDLE start. If we get demand for the old
behavior, we can consider re-adding it, but I bet when they hear the
downside (a second gettimeofday() call for every query), no one will
want the old behavior.

---------------------------------------------------------------------------

Tom Lane wrote:
> I redid my previous measurements after finishing up the weekend's
> hacking. The numbers shown below are elapsed time in seconds for
>
> time psql -f testfile.sql postgres >/dev/null
>
> using CVS HEAD and REL8_1_STABLE branch tip, compiled --enable-debug
> --disable-cassert, no nondefault options except for turning fsync off
> (which doesn't particularly affect read-only tests like these anyway).
> The machines are both running current Fedora Core 5. The older x86
> machine is known to have the slow-gettimeofday problem from previous
> experimentation with EXPLAIN ANALYZE. Each number is the median of 3
> or more tests, rounded off to 0.1 second (I wouldn't put a lot of faith
> in differences of 0.1 sec or so, because of the variance I saw in the
> tests).
>
> x86 x86_64
>
> 8.1 HEAD 8.1 HEAD
>
> 100000 "SELECT 1;" 25.9 27.0 9.2 9.1
> with stats_command_string=1 63.5 27.6 18.7 9.2
> with log_min_duration_statement=100 26.9 27.8 9.6 9.2
> with statement_timeout=100 27.5 28.6 9.6 9.8
> with all 3 features 66.1 29.3 19.5 9.7
>
> BEGIN, 100000 "SELECT 1;", COMMIT 21.2 23.1 8.3 8.4
> with stats_command_string=1 52.3 23.5 15.4 8.5
> with log_min_duration_statement=100 22.1 23.4 8.4 8.4
> with statement_timeout=100 23.7 24.3 8.6 8.8
> with all 3 features 55.2 25.5 16.0 8.8
>
> I chose the log_min_duration_statement and statement_timeout settings
> high enough so that no actual logging or timeout would happen --- the
> point is to measure the measurement overhead.
>
> The good news is that we've pretty much licked the problem of
> stats_command_string costing an unreasonable amount.
>
> The bad news is that except in the stats_command_string cases, HEAD
> is noticeably slower than 8.1 on the machine with slow gettimeofday.
> In the single-transaction test this might be blamed on the addition
> of statement_timestamp support (which requires a gettimeofday per
> statement that wasn't there in 8.1) ... but in the one-transaction-
> per-statement tests that doesn't hold water, because each branch is
> doing a gettimeofday per statement, just in different places.
>
> Can anyone else reproduce this slowdown? It might be only an artifact
> of these particular builds, but it's a bit too consistent in my x86 data
> to just ignore.
>
> BTW, according to "top" the CPU usage percentages in these tests are
> on the order of 55% backend, 45% psql. Methinks psql needs a round
> of performance tuning ...
>
> regards, tom lane
>
> ---------------------------(end of broadcast)---------------------------
> TIP 6: explain analyze is your friend
>

--
Bruce Momjian bruce@momjian.us
EnterpriseDB http://www.enterprisedb.com

+ If your life is a hard drive, Christ can be your backup. +

---------------------------(end of broadcast)---------------------------
TIP 4: Have you searched our list archives?

http://archives.postgresql.org

Tom Lane wrote:

> The bad news is that except in the stats_command_string cases, HEAD
> is noticeably slower than 8.1 on the machine with slow gettimeofday.
> In the single-transaction test this might be blamed on the addition
> of statement_timestamp support (which requires a gettimeofday per
> statement that wasn't there in 8.1) ... but in the one-transaction-
> per-statement tests that doesn't hold water, because each branch is
> doing a gettimeofday per statement, just in different places.
>
> Can anyone else reproduce this slowdown? It might be only an artifact
> of these particular builds, but it's a bit too consistent in my x86 data
> to just ignore.

This is what I get on a fast AMD Dual Opteron box(Running Debian
Sarge/AMD64):

8.1.4 HEAD
1000000 SELECT 1; 74,74,73 77,76,77
stats_command_string=1; 105,99,106 78,79,78
log_min_duration_statement=100 79,80,81 75,80,76
statement_timeout=100 78,79,78 75,79,77
all 3 104,108,107 82,81,81

all values in seconds with 3 consecutive runs of one million "SELECT 1;"
queries. It takes about 48 seconds to run the same test without
stat-collection btw.


Stefan

---------------------------(end of broadcast)---------------------------
TIP 6: explain analyze is your friend

Stefan Kaltenbrunner <stefan@kaltenbrunner.cc> writes:
> This is what I get on a fast AMD Dual Opteron box(Running Debian
> Sarge/AMD64):

> 8.1.4 HEAD
> 1000000 SELECT 1; 74,74,73 77,76,77
> stats_command_string=1; 105,99,106 78,79,78
> log_min_duration_statement=100 79,80,81 75,80,76
> statement_timeout=100 78,79,78 75,79,77
> all 3 104,108,107 82,81,81

> all values in seconds with 3 consecutive runs of one million "SELECT 1;"
> queries. It takes about 48 seconds to run the same test without
> stat-collection btw.

I'm confused. Isn't your first table row for the case of no stat collection?
Or do you mean that you have stats_row_level and/or stats_block_level on
in all four cases?

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 6: explain analyze is your friend

Tom Lane wrote:
> Stefan Kaltenbrunner <stefan@kaltenbrunner.cc> writes:
>> This is what I get on a fast AMD Dual Opteron box(Running Debian
>> Sarge/AMD64):
>
>> 8.1.4 HEAD
>> 1000000 SELECT 1; 74,74,73 77,76,77
>> stats_command_string=1; 105,99,106 78,79,78
>> log_min_duration_statement=100 79,80,81 75,80,76
>> statement_timeout=100 78,79,78 75,79,77
>> all 3 104,108,107 82,81,81
>
>> all values in seconds with 3 consecutive runs of one million "SELECT 1;"
>> queries. It takes about 48 seconds to run the same test without
>> stat-collection btw.
>
> I'm confused. Isn't your first table row for the case of no stat collection?
> Or do you mean that you have stats_row_level and/or stats_block_level on
> in all four cases?

yes - stats_row_level and stats_block_level on in all cases (sorry for
the confusion) - I can easily redo the tests without those - but that's
what I had in the running conf and I only remember that after I was
nearly done with all the testing :-)



Stefan

---------------------------(end of broadcast)---------------------------
TIP 3: Have you checked our extensive FAQ?

http://www.postgresql.org/docs/faq

Stefan Kaltenbrunner <stefan@kaltenbrunner.cc> writes:
> Tom Lane wrote:
>> Or do you mean that you have stats_row_level and/or stats_block_level on
>> in all four cases?

> yes - stats_row_level and stats_block_level on in all cases (sorry for
> the confusion) - I can easily redo the tests without those - but that's
> what I had in the running conf and I only remember that after I was
> nearly done with all the testing :-)

It'd be interesting to compare 8.1 and HEAD for the no-overhead case;
I don't think you need to redo all four cases, but I'd like to see that one.

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 6: explain analyze is your friend

Tom Lane wrote:
> Stefan Kaltenbrunner <stefan@kaltenbrunner.cc> writes:
>> Tom Lane wrote:
>>> Or do you mean that you have stats_row_level and/or stats_block_level on
>>> in all four cases?
>
>> yes - stats_row_level and stats_block_level on in all cases (sorry for
>> the confusion) - I can easily redo the tests without those - but that's
>> what I had in the running conf and I only remember that after I was
>> nearly done with all the testing :-)
>
> It'd be interesting to compare 8.1 and HEAD for the no-overhead case;
> I don't think you need to redo all four cases, but I'd like to see that one.

8.1: 50,50,49
HEAD: 49,48,49


Stefan

---------------------------(end of broadcast)---------------------------
TIP 9: In versions below 8.0, the planner will ignore your desire to
choose an index scan if your joining column's datatypes do not
match

Stefan Kaltenbrunner <stefan@kaltenbrunner.cc> writes:
> Tom Lane wrote:
>> It'd be interesting to compare 8.1 and HEAD for the no-overhead case;
>> I don't think you need to redo all four cases, but I'd like to see that one.

> 8.1: 50,50,49
> HEAD: 49,48,49

OK, so that seems comparable to my results on a dual Xeon ... probably,
both your machine and my newer one have fast-to-read clock hardware.
We need to get some numbers from one of the people who have complained
about EXPLAIN ANALYZE overhead.

I'll have to try the stats-collection-active case on my machines, too.
I was still planning to look into removing the buffer process to reduce
context-swap overhead for stats collection ...

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 2: Don't 'kill -9' the postmaster

Tom Lane wrote:
> Stefan Kaltenbrunner <stefan@kaltenbrunner.cc> writes:
> > Tom Lane wrote:
> >> It'd be interesting to compare 8.1 and HEAD for the no-overhead case;
> >> I don't think you need to redo all four cases, but I'd like to see that one.
>
> > 8.1: 50,50,49
> > HEAD: 49,48,49
>
> OK, so that seems comparable to my results on a dual Xeon ... probably,
> both your machine and my newer one have fast-to-read clock hardware.
> We need to get some numbers from one of the people who have complained
> about EXPLAIN ANALYZE overhead.

I'm compiling here without the assert stuff. It takes a while ...

--
Alvaro Herrera http://www.CommandPrompt.com/
The PostgreSQL Company - Command Prompt, Inc.

---------------------------(end of broadcast)---------------------------
TIP 5: don't forget to increase your free space map settings