performance issue with a 2.5gb joinded table

[Daniel Westermann]

Hi Listers,

we migrated an oracle datawarehouse to postgresql 9.1 ( ppas 9.1.7.12 ) and are facing massive issues with response times in postgres when compared to the oracle system. Both database run on the same hardware and storage ( rhel5.8 64bit ).

Oracle memory parameters are:
SGA=1gb
PGA=200mb

Postgres currently runs with 15gb of shared buffers ( that’s because the big table in question is around 2.5gb in size and one suggestion was to increase that much so postgresql will cache the complete table. and this is the case now ).

explain (analyze,buffers) SELECT test1.slsales_batch
, test1.slsales_checksum
, test1.slsales_reg_id
, test1.slsales_prod_id
, test1.slsales_date_id
, test1.slsales_pos_id
, test1.slsales_amt_sales_gross
, test1.slsales_amt_sales_discount
, test1.slsales_units_sales_gross
, test1.slsales_amt_returns
, test1.slsales_amt_returns_discount
, test1.slsales_units_returns
, (test1.slsales_amt_sales_gross - test1.slsales_amt_returns)
* mgmt_fact_winratio.winratio_ratio AS slsales_amt_est_winnings
, mgmt_fact_winratio.winratio_ratio AS slsales_ratio
FROM mgmtt_own.test1
LEFT JOIN mgmtt_own.mgmt_fact_winratio
ON mgmt_fact_winratio.winratio_date_id = test1.slsales_date_id

Oracle’s explain plan looks like this:

----------------------------------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes |TempSpc| Cost (%CPU)| Time |
----------------------------------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 25M| 1527M| | 115K (3)| 00:23:10 |
|* 1 | HASH JOIN RIGHT OUTER| | 25M| 1527M| 4376K| 115K (3)| 00:23:10 |
| 2 | TABLE ACCESS FULL | MGMT_FACT_WINRATIO | 159K| 2498K| | 167 (5)| 00:00:03 |
| 3 | TABLE ACCESS FULL | TEST1 | 25M| 1139M| | 43435 (5)| 00:08:42 |
----------------------------------------------------------------------------------------------------
Predicate Information (identified by operation id):
---------------------------------------------------
1 - access("MGMT_FACT_WINRATIO"."WINRATIO_PROD_ID"(+)="TEST1"."SLSALES_PROD_ID" AND
"MGMT_FACT_WINRATIO"."WINRATIO_DATE_ID"(+)="TEST1"."SLSALES_DATE_ID")

Somehow oracle seems to know that a right join is the better way to go.

Postgres’s explain plan:

QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------
Hash Left Join (cost=3948.52..13646089.21 rows=25262160 width=61) (actual time=260.642..81240.692 rows=25262549 loops=1)
Hash Cond: ((test1.slsales_date_id = mgmt_fact_winratio.winratio_date_id) AND (test1.slsales_prod_id = mgmt_fact_winratio.winratio_prod_id))
Buffers: shared hit=306590
-> Seq Scan on test1 (cost=0.00..254148.75 rows=25262160 width=56) (actual time=0.009..15674.535 rows=25262161 loops=1)
Buffers: shared hit=305430
-> Hash (cost=1582.89..1582.89 rows=157709 width=19) (actual time=260.564..260.564 rows=157709 loops=1)
Buckets: 16384 Batches: 1 Memory Usage: 7855kB
Buffers: shared hit=1160
-> Seq Scan on mgmt_fact_winratio (cost=0.00..1582.89 rows=157709 width=19) (actual time=0.008..114.406 rows=157709 loops=1)
Buffers: shared hit=1160
Total runtime: 95762.025 ms
(11 rows)

Tried to modify the statement according to oracle’s plan, but this did not help:

explain (analyze,buffers) SELECT test1.slsales_batch
, test1.slsales_checksum
, test1.slsales_reg_id
, test1.slsales_prod_id
, test1.slsales_date_id
, test1.slsales_pos_id
, test1.slsales_amt_sales_gross
, test1.slsales_amt_sales_discount
, test1.slsales_units_sales_gross
, test1.slsales_amt_returns
, test1.slsales_amt_returns_discount
, test1.slsales_units_returns
, (test1.slsales_amt_sales_gross - test1.slsales_amt_returns)
* mgmt_fact_winratio.winratio_ratio AS slsales_amt_est_winnings
, mgmt_fact_winratio.winratio_ratio AS slsales_ratio
FROM mgmtt_own.test1
, mgmtt_own.mgmt_fact_winratio
WHERE mgmt_fact_winratio.winratio_prod_id(+) = test1.slsales_prod_id
AND mgmt_fact_winratio.winratio_date_id(+) = test1.slsales_date_id
;
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------------------------------
Hash Left Join (cost=3948.52..13646089.21 rows=25262160 width=61) (actual time=276.605..80629.400 rows=25262549 loops=1)
Hash Cond: ((test1.slsales_prod_id = mgmt_fact_winratio.winratio_prod_id) AND (test1.slsales_date_id = mgmt_fact_winratio.winratio_date_id))
Buffers: shared hit=306590
-> Seq Scan on test1 (cost=0.00..254148.75 rows=25262160 width=56) (actual time=0.009..15495.167 rows=25262161 loops=1)
Buffers: shared hit=305430
-> Hash (cost=1582.89..1582.89 rows=157709 width=19) (actual time=276.515..276.515 rows=157709 loops=1)
Buckets: 16384 Batches: 1 Memory Usage: 7855kB
Buffers: shared hit=1160
-> Seq Scan on mgmt_fact_winratio (cost=0.00..1582.89 rows=157709 width=19) (actual time=0.009..119.930 rows=157709 loops=1)
Buffers: shared hit=1160
Total runtime: 95011.401 ms

Parameters changed:
default_statistics_target =1000
enable_mergejoin=false ( when enabled query takes even longer )
seq_page_cost=1
random_page_cost=2

vacuumed the whole database and currently there is no data coming in, so everything is up to date.

What additionally makes me wonder is, that the same table in oracle is taking much less space than in postgresql:

SQL> select sum(bytes) from dba_extents where segment_name = 'TEST1';
SUM(BYTES)
----------
1610612736

select pg_relation_size('mgmtt_own.test1');
pg_relation_size
------------------
2502082560
(1 row)

(sysdba@[local]:7777) [bi_dwht] > \d+ mgmtt_own.test1
Table "mgmtt_own.test1"
Column | Type | Modifiers | Storage | Description
------------------------------+---------------+-----------+---------+-------------
slsales_batch | numeric(8,0) | | main |
slsales_checksum | numeric(8,0) | | main |
slsales_reg_id | numeric(8,0) | | main |
slsales_prod_id | numeric(8,0) | | main |
slsales_date_id | numeric(8,0) | | main |
slsales_pos_id | numeric(8,0) | | main |
slsales_amt_sales_gross | numeric(16,6) | | main |
slsales_amt_sales_discount | numeric(16,6) | | main |
slsales_units_sales_gross | numeric(8,0) | | main |
slsales_amt_returns | numeric(16,6) | | main |
slsales_amt_returns_discount | numeric(16,6) | | main |
slsales_units_returns | numeric(8,0) | | main |
slsales_amt_est_winnings | numeric(16,6) | | main |
Indexes:
"itest1" btree (slsales_date_id) CLUSTER, tablespace "mgmtt_idx"
"itest2" btree (slsales_prod_id), tablespace "mgmtt_idx"
Has OIDs: no
Tablespace: "mgmtt_dat"

Although the plan seems to be ok because most of the table must be read 95 secs compared to 23 secs will be a killer for the project.

Any hints what else could be checked/done ?

Kind Regards
Daniel

[Heikki Linnakangas]

SQL> select sum(bytes) from dba_extents where segment_name = 'TEST1';
SUM(BYTES)
----------
1610612736
select pg_relation_size('mgmtt_own.test1');
pg_relation_size
------------------
2502082560
(1 row)
Table "mgmtt_own.test1"
Column | Type | Modifiers | Storage | Description
------------------------------+---------------+-----------+---------+-------------
slsales_batch | numeric(8,0) | | main |
slsales_checksum | numeric(8,0) | | main |
slsales_reg_id | numeric(8,0) | | main |
slsales_prod_id | numeric(8,0) | | main |
slsales_date_id | numeric(8,0) | | main |
slsales_pos_id | numeric(8,0) | | main |
slsales_amt_sales_gross | numeric(16,6) | | main |
slsales_amt_sales_discount | numeric(16,6) | | main |
slsales_units_sales_gross | numeric(8,0) | | main |
slsales_amt_returns | numeric(16,6) | | main |
slsales_amt_returns_discount | numeric(16,6) | | main |
slsales_units_returns | numeric(8,0) | | main |
slsales_amt_est_winnings | numeric(16,6) | | main |
"itest1" btree (slsales_date_id) CLUSTER, tablespace "mgmtt_idx"
"itest2" btree (slsales_prod_id), tablespace "mgmtt_idx"
Has OIDs: no
Tablespace: "mgmtt_dat"

One difference is that numerics are stored more tightly packed on
Oracle. Which is particularly good for Oracle as they don't have other
numeric data types than number. On PostgreSQL, you'll want to use int4
for ID-fields, where possible. An int4 always takes up 4 bytes, while a
numeric holding an integer value in the same range is typically 5-9 bytes.

- Heikki

[Daniel Westermann]

-----Original Message-----
From: Heikki Linnakangas [mailto:***@vmware.com]
Sent: Donnerstag, 3. Januar 2013 18:02
To: Daniel Westermann
Cc: 'pgsql-***@postgresql.org'
Subject: Re: [PERFORM] FW: performance issue with a 2.5gb joinded table

SQL> select sum(bytes) from dba_extents where segment_name =
SQL> 'TEST1';
SUM(BYTES)
----------
1610612736
select pg_relation_size('mgmtt_own.test1');
pg_relation_size
------------------
2502082560
(1 row)
Table "mgmtt_own.test1"
Column | Type | Modifiers | Storage | Description
------------------------------+---------------+-----------+---------+-
------------------------------+---------------+-----------+---------+-
------------------------------+---------------+-----------+---------+-
------------------------------+---------------+-----------+---------+-
------------------------------+---------------+-----------+---------+-
------------------------------+---------------+-----------+---------+-
------------------------------+---------------+-----------+---------+-
------------------------------+---------------+-----------+---------+-
------------------------------+---------------+-----------+---------+-
------------------------------+---------------+-----------+---------+-
------------------------------+---------------+-----------+---------+-
------------------------------+---------------+-----------+---------+-
------------------------------+---------------+-----------+---------+-
slsales_batch | numeric(8,0) | | main |
slsales_checksum | numeric(8,0) | | main |
slsales_reg_id | numeric(8,0) | | main |
slsales_prod_id | numeric(8,0) | | main |
slsales_date_id | numeric(8,0) | | main |
slsales_pos_id | numeric(8,0) | | main |
slsales_amt_sales_gross | numeric(16,6) | | main |
slsales_amt_sales_discount | numeric(16,6) | | main |
slsales_units_sales_gross | numeric(8,0) | | main |
slsales_amt_returns | numeric(16,6) | | main |
slsales_amt_returns_discount | numeric(16,6) | | main |
slsales_units_returns | numeric(8,0) | | main |
slsales_amt_est_winnings | numeric(16,6) | | main |
"itest1" btree (slsales_date_id) CLUSTER, tablespace "mgmtt_idx"
"itest2" btree (slsales_prod_id), tablespace "mgmtt_idx"
Has OIDs: no
Tablespace: "mgmtt_dat"

One difference is that numerics are stored more tightly packed on Oracle. Which is particularly good for Oracle as they don't have other numeric data types than number. On PostgreSQL, you'll want to use int4 for ID-fields, where possible. An int4 always takes up 4 bytes, while a numeric holding an integer value in the same range is typically 5-9 bytes.

- Heikki

Thanks for poiting that out, Heikki.

[Tom Lane]

Post by Heikki Linnakangas
One difference is that numerics are stored more tightly packed on
Oracle. Which is particularly good for Oracle as they don't have other
numeric data types than number. On PostgreSQL, you'll want to use int4
for ID-fields, where possible. An int4 always takes up 4 bytes, while a
numeric holding an integer value in the same range is typically 5-9 bytes.

Replacing those numeric(8) and numeric(16) fields with int4 and int8
would be greatly beneficial to comparison and hashing performance,
not just table size. I'm a bit surprised that EDB's porting tools
evidently don't do this automatically (I infer from the reference to
PPAS that the OP is using EDB ...)

regards, tom lane

[Daniel Westermann]

-----Original Message-----
From: Tom Lane [mailto:***@sss.pgh.pa.us]
Sent: Freitag, 4. Januar 2013 21:41
To: Heikki Linnakangas
Cc: Daniel Westermann; 'pgsql-***@postgresql.org'
Subject: Re: [PERFORM] FW: performance issue with a 2.5gb joinded table

Post by Heikki Linnakangas
One difference is that numerics are stored more tightly packed on
Oracle. Which is particularly good for Oracle as they don't have other
numeric data types than number. On PostgreSQL, you'll want to use int4
for ID-fields, where possible. An int4 always takes up 4 bytes, while
a numeric holding an integer value in the same range is typically 5-9 bytes.

Replacing those numeric(8) and numeric(16) fields with int4 and int8 would be greatly beneficial to comparison and hashing performance, not just table size. I'm a >> bit surprised that EDB's porting tools evidently don't do this automatically (I infer from the reference to PPAS that the OP is using EDB ...)
regards, tom lane

Thanks, tom. Any clue where there remaining around 500mb difference come from ? converted all the numeric(8) to int and this saved around 380mb of storage and around 10 secs exectution time... both databases have their files on standard ext3, same fs options. Given that the table has around 25'000'000 rows this is still approx. 20 bytes more per row on average

Regards
Daniel

[Vladimir Sitnikov]

Daniel,

Post by Daniel Westermann
Somehow oracle seems to know that a right join is the better way to go.

In fact, PostgreSQL is just doing the same thing: it hashes smaller table
and scans the bigger one.

Could you please clarify how do you consume 25M rows?
It could be the difference of response times comes not from the PostgreSQL
itself, but from the client code.

Could you please add the following information?
1) Execution time of simple query that selects MAX of all the required
columns "select max(test1.slsales_batch) , max(test1.slsales_checksum),
...".
I mean not explain (analyze, buffers), but simple execution.
The purpose of MAX is to split overhead of consuming of the resultset from
the overhead of producing it.

2) explain (analyze, buffers) for the same query with maxes. That should
reveal the overhead of explain analyze itself.

3) The output of the following SQLPlus script (from Oracle):
set linesize 1000 pagesize 10000 trimout on trimspool on time on timing on
spool slow_query.lst
select /*+ gather_plan_statistics */ max(test1.slsales_batch) ,
max(test1.slsales_checksum), ..;
select * from table(dbms_xplan.display_cursor(null, null, 'ALLSTATS
LAST'));
spool off

That would display detailed statistics on execution time similar to the
explain (analyze, buffers).

4) Could you please clarify how did you migrate test1 table?
I guess the order of rows in that table might affect overall execution time.
Sorted table would be more CPU cache friendly, thus giving speedup. (see
[1] for similar example).
As far as I understand, simple create table as select * from test1 order by
slsales_date_id, slsales_prod_id should improve cache locality.


[1]:
http://stackoverflow.com/questions/11227809/why-is-processing-a-sorted-array-faster-than-an-unsorted-array