Hello

I need high throughput while inserting into PostgreSQL. Because of that I
did some PostgreSQL insert performance tests.

------------------------------------------------------------
-- Test schema
create table logs (
logid serial primary key,
ctime integer not null,
stime integer not null,
itime integer not null,
agentid integer not null,
subagentid integer not null,
ownerid integer not null,
hostid integer not null,
appname varchar(64) default null,
logbody varchar(1024) not null
);

create index ctime_ndx on logs using btree (ctime);
create index stime_ndx on logs using btree (stime);
create index itime_ndx on logs using btree (itime);
create index agentid_ndx on logs using hash (agentid);
create index ownerid_ndx on logs using hash (ownerid);
create index hostid_ndx on logs using hash (hostid);
------------------------------------------------------------

Test Hardware:
IBM Thinkpad R40
CPU: Pentium 4 Mobile 1993 Mhz (full powered)
RAM: 512 MB
OS: GNU/Linux, Fedora Core 1, kernel 2.4.24

A test program developed with libpq inserts 200.000 rows into table
logs. Insertions are made with 100 row per transaction (total 2.000
transactions).

Some parameter changes from postgresql.conf file follows:
----------------------------------------------------------------
shared_buffers = 2048 # min max_connections*2 or 16, 8KB each
max_fsm_relations = 20000 # min 10, fsm is free space map, ~40 bytes
max_fsm_pages = 200000 # min 1000, fsm is free space map, ~6 bytes
max_locks_per_transaction = 256 # min 10
wal_buffers = 64 # min 4, typically 8KB each
sort_mem = 32768 # min 64, size in KB
vacuum_mem = 16384 # min 1024, size in KB
checkpoint_segments = 6 # in logfile segments, min 1, 16MB each
checkpoint_timeout = 900 # range 30-3600, in seconds
fsync = true
wal_sync_method = fsync # the default varies across platforms:
enable_seqscan = true
enable_indexscan = true
enable_tidscan = true
enable_sort = true
enable_nestloop = true
enable_mergejoin = true
enable_hashjoin = true
effective_cache_size = 2000 # typically 8KB each
geqo = true
geqo_selection_bias = 2.0 # range 1.5-2.0
geqo_threshold = 11
geqo_pool_size = 0 # default based on tables in statement,
# range 128-1024
geqo_effort = 1
geqo_generations = 0
geqo_random_seed = -1 # auto-compute seed
----------------------------------------------------------------

The test was made with both of PostgreSQL 7.3.4 and PostgreSQL 7.4.1 (the
test program was recompiled during version changes).

The results are below (average inserted rows per second).

speed for speed for
# of EXISTING RECORDS PostgreSQL 7.3.4 PostgreSQL 7.4.1
================================================== =======================

0 initial records 1086 rows/s 1324 rows/s
200.000 initial records 781 rows/s 893 rows/s
400.000 initial records 576 rows/s 213 rows/s
600.000 initial records 419 rows/s 200 rows/s
800.000 initial records 408 rows/s not tested because of bad
results


When the logs table reconstructed with only one index (primary key) then
2941 rows/s speed is reached. But I need all the seven indexes.

The question is why the PostgreSQL 7.4.1 is so slow under heavy work?

Is there a way to speed up inserts without eliminating indexes?

What about concurrent inserts (cocurrent spare test program execution)
into the same table? It did not work.

-sezai

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

Sezai YILMAZ wrote:
I suggest you up that to say 10000 buffers..

You need to pull it down a little, I guess. How about 8/16MB?

Not required. 1024 could be done since you are testing inserts anyways. Of
course, it matters only when you run vacuum..

Is that true? It tells postgresql that it has around 16MB memory. Set it up
around 15000 so that around 100MB+ is used. Might change the results of index
scans.. I always prefer to set it to whatever available.

Do you mean 800000? I believe the '.' is a thousands separator here but not too
sure..:-)

Can you run vmstat and see where things get stalled? Probably you can up the
number of WAL segments and attempt.

What does it mean, it didn't work? Any errors?

HTH

Shridhar



---------------------------(end of broadcast)---------------------------
TIP 1: subscribe and unsubscribe commands go to majordomo (AT) postgresql (DOT) org

I don't know the answer to the question of why 7.4 is slower, but I have
some suggestions on additional things to test, and how to make it faster.

First off, try 200 transactions of 1000 records each, you might even want
to try 20 transactions of 10,000 records each. Postgres seems to run much
faster the less commits you have, but different configs may change the
sweet spot.

Secondly, one possible solution to your problem is to drop the indexes,
insert the new rows and recreate the indexes. Of course, for testing,
you'll want to time the entire process of drop/insert/create and compare
it to the raw insert time with indexes intact. I use a stored procedure
on my databases, i.e.:

select drop_foo_indexes();
....
<commands to insert many rows into table foo>
....
select create_foo_indexes();

Another thing to consider is vacuums. You don't mention how often you
vacuumed the database during testing, I would recommend a "vacuum full"
between each test (unless, of course, you're testing how much a lack
of vacuum hurts performance

Hope this helps.

Sezai YILMAZ wrote:
--
Bill Moran
Potential Technologies
http://www.potentialtech.com


---------------------------(end of broadcast)---------------------------
TIP 3: if posting/reading through Usenet, please send an appropriate
subscribe-nomail command to majordomo (AT) postgresql (DOT) org so that your
message can get through to the mailing list cleanly

Sezai YILMAZ wrote:

I changed the three hash indexes to btree.

The performance is increased about 2 times (in PostgreSQL 7.3.4 1905
rows/s).

Concurrent inserts now work.

Changed indexes are more suitable for hash type. Because, there is no
ordering on them, instead exact values are matched which is more natural
for hash type of indexes. But hash indexes has possible dead lock
problems on multiple concurrent inserts. I think I can live with btree
indexes. They work better. :-)

-sezai

---------------------------(end of broadcast)---------------------------
TIP 6: Have you searched our list archives?

http://archives.postgresql.org

Hash indexes have relatively poor concurrency, though I think it should still
work. You probably want to be using btree indexes for everything though,
unless you can actually profile the two and show hash indexes being a big win.

Note that there were bugs in the hash index code at least through most 7.3
versions.

--
greg


---------------------------(end of broadcast)---------------------------
TIP 4: Don't 'kill -9' the postmaster

Sezai YILMAZ <sezai.yilmaz (AT) pro-g (DOT) com.tr> writes:
Concurrent inserts should work with hash indexes in 7.4, though not 7.3.

The slowdown you report probably is due to the rewrite of hash indexing
to allow more concurrency --- the locking algorithm is more complex than
it used to be. I am surprised that the effect is so large though.
Could you make your test program available?

Are they? How many distinct values are there in those columns?
I suspect that your test may be stressing the case where only a few hash
buckets are used and each bucket chain gets to be very long.

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 1: subscribe and unsubscribe commands go to majordomo (AT) postgresql (DOT) org

Tom Lane wrote:

hash indexes. 7.4 solves this problem but is very slow.


Comiple and link scenarios:

without transactions (where each insert is a transaction)
$ gcc -o tester tester.c -lpq

with default 400 inserts per transaction blocks
$ gcc -DTRANSACTION -o tester tester.c -lpq

with 200 inserts per transaction blocks
$ gcc -DTRANSACTION -DINSERTPERTRANSACTION=200 -o tester tester.c -lpq

I do concurrent tests by starting seperate tester programs from
different xterm windows.

information is in "tester.c" where INSERT query string is built.

Regards,

-sezai


---------------------------(end of broadcast)---------------------------
TIP 3: if posting/reading through Usenet, please send an appropriate
subscribe-nomail command to majordomo (AT) postgresql (DOT) org so that your
message can get through to the mailing list cleanly

On Saturday 28 February 2004 13:59, Sezai YILMAZ wrote:
Some tests on CVS head in case somebody finds the data interesting. It is a
single IDE disk system with linux 2.6.2 running. It has 512MB RAM and 2.66GHz
P-IV. The file system is reiserfs.

I pulled CVS head couple of days back.

Everything default except for shared_buffers=100 and effective cache=25000, I
got 1980 inserts/sec in a single run.

With checkpoint segments 10, I got 1923 inserts per sec.

With two concurrent processes and 10 checkpoint segments, I got 1673 req/sec.

I noted that in vmstat, the IO wasn't pushed really hard. The block out were
varying about 1000-5000 per sec. However occasionally that would spike to
18000 blocks. I guess that would be some checkpoint going on.

and I could not find sort_mem in postgresql.conf. Is work_mem new name for it?
I recall the discussion to sanitize the name but not the result of it..

Shridhar

---------------------------(end of broadcast)---------------------------
TIP 6: Have you searched our list archives?

http://archives.postgresql.org

Shridhar Daithankar <shridhar (AT) frodo (DOT) hserus.net> writes:
100?

Yeah.

regards, tom lane

---------------------------(end of broadcast)---------------------------
TIP 6: Have you searched our list archives?

http://archives.postgresql.org

On Saturday 28 February 2004 21:27, Tom Lane wrote:
1000.. That was a typo..

Shridhar

---------------------------(end of broadcast)---------------------------
TIP 3: if posting/reading through Usenet, please send an appropriate
subscribe-nomail command to majordomo (AT) postgresql (DOT) org so that your
message can get through to the mailing list cleanly