[ADMIN] Fwd: My Query to insert and retrieve takes time

[Shivakumar Ramannavar]

Hi Community,

We are analyzing performance behaviour with postgres database, for INSERT
operation for 3,00,000 records (On SLES 10 Linux, 64 bit server) postgreSQL
is taking around 150 min.

Here is some observation drawn from PostgreSQL performance testing:

There are around 310,000 records, it takes 8.313 milliseconds to
retrieval/search a record, and it takes 8.321 milliseconds more to update a
record. Totally it takes approximate of 17.9289 ms for searching and update
during a Synchronisation operation (these timings are recorded from
postgres Embedded SQL modules).

This is very slow compared to Sybase for the same INSERT operation (with
3,00,000 students). There are around 300,000 update operations and it is
taking approx 150 min as against 30 - 40 min at Sybase (when run the same
operation/hardware).

However when checked at PSQL client, the queries are pretty faster, and the
below is the explain PLANS of SELECT and UPDATE ((student_id is UNIQUE
index, and id is primary key) :

EXPLAIN ANALYSE SELECT id from students where student_id='9101';
QUERY PLAN
------------------------------------------------------------------------------------------------------------------------
Index Scan using student_ind2 on students(cost=0.00..8.38 rows=1 width=4)
(actual time=0.096..0.097 rows=1 loops=1)
Index Cond: (student_id = 9101)
*Total runtime: 0.111 ms*
*
*
(3 rows)


EXPLAIN ANALYZE UPDATE students SET marks = 10
WHERE student_id = '9101'

QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------
Update (cost=0.00..8.38 rows=1 width=64) (actual time=0.039..0.039 rows=0
loops=1)
-> Index Scan using student_ind2 on students (cost=0.00..8.38 rows=1
width=64) (actual time=0.016..0.017 rows=1 loops=1)
Index Cond: (student_id = 9101)

* Total runtime: 0.080 ms*

As per the above plans, the queries are using indexes as per our
expectation.

Can anyone help us in improving the performance,
Shiva

---------- Forwarded message ----------
From: Shivakumar Ramannavar <***@gmail.com>
Date: Mon, Oct 31, 2011 at 4:51 PM
Subject: My Query to insert and retrieve takes time
To: pgsql-***@postgresql.org


Hi Community,

Require some help, I am running postgres 9.1 database, which has table
having around 300,000 records, my problem is that it takes 8.314
milliseconds to update one record, and the below is the explain PLAN of
(terminal_id is UNIQUE index, and term_index is primary key)

EXPLAIN ANALYZE UPDATE terminal
SET unit_address=E'\\x00000015a3'
WHERE terminal_id = '320000'


QUERY PLAN
-------------------------------------------------------------------------------------------------------------------------------
Update (cost=0.00..8.38 rows=1 width=64) (actual time=0.074..0.074 rows=0
loops=1)
-> Index Scan using terminal_ind2 on terminal (cost=0.00..8.38 rows=1
width=64) (actual time=0.047..0.048 rows=1 loops=1)
Index Cond: (terminal_id = 320000)
Total runtime: 0.182 ms

There are around 300,000 update operations and it is taking approx 80 min,
please let me know how to improve the performance.

Thanks!

[Tom Lane]

Post by Shivakumar Ramannavar
There are around 310,000 records, it takes 8.313 milliseconds to
retrieval/search a record, and it takes 8.321 milliseconds more to update a
record.

It was already pointed out to you that these numbers are suspiciously
close together, and also suspiciously close to the rotation time of a
7200RPM disk. It seems very likely that you have configured things so
that each operation is done in a transaction that has to be explicitly
committed to disk, thus requiring a write to WAL, which would limit the
transaction rate to one per rotation. However, since you haven't shown
us anything about how the queries are being issued, we can't confirm
or disprove that theory.

Consider batching multiple operations into a single transaction.
Turning off synchronous_commit is another possibility, depending on
what your crash-safety requirements are.

regards, tom lane