Why is PostgreSQL 9.2 slower than 9.1 in my tests? - postgresql

Question

I have upgraded from PostgreSQL 9.1.5 to 9.2.1:
"PostgreSQL 9.1.5 on x86_64-unknown-linux-gnu, compiled by gcc (GCC) 4.4.6 20120305 (Red Hat 4.4.6-4), 64-bit"
"PostgreSQL 9.2.1 on x86_64-unknown-linux-gnu, compiled by gcc (GCC) 4.4.6 20120305 (Red Hat 4.4.6-4), 64-bit"
It is on the same machine with default PostgreSQL configuration files (only port was changed).
For testing purpose I have simple table:
CREATE TEMP TABLE test_table_md_speed(id serial primary key, n integer);
Which I test using function:
CREATE OR REPLACE FUNCTION TEST_DB_SPEED(cnt integer) RETURNS text AS $$
DECLARE
time_start timestamp;
time_stop timestamp;
time_total interval;
BEGIN
time_start := cast(timeofday() AS TIMESTAMP);
FOR i IN 1..cnt LOOP
INSERT INTO test_table_md_speed(n) VALUES (i);
END LOOP;
time_stop := cast(timeofday() AS TIMESTAMP);
time_total := time_stop-time_start;
RETURN extract (milliseconds from time_total);
END;
$$ LANGUAGE plpgsql;
And I call:
SELECT test_db_speed(1000000);
I see strange results. For PostgreSQL 9.1.5 I get "8254.769", and for 9.2.1 I get: "9022.219". This means that new version is slower. I cannot find why.
Any ideas why those results differ?

Answer 1

You say both are on the same machine. Presumably the data files for the newer version were added later. Later files tend to be added closer to the center of the platter, where access speeds are slower.
There is a good section on this in Greg Smith's book on PostgreSQL performance, including ways to measure and graph the effect. With clever use of the dd utility you might be able to do some ad hoc tests of the relative speed at each location, at least for reads.
The 9.2 release generally scales up to a large number of cores better than earlier versions, although in some of the benchmarks there was a very slight reduction in the performance of a single query running alone. I didn't see any benchmarks showing an effect anywhere near this big, though; I would bet on it being the result of position on the drive -- with just goes to show how hard it can be to do good benchmarking.
UPDATE: A change made in 9.2.0 to improve performance for some queries made some other queries perform worse. Eventually it was determined that this change should be reverted, which happened in version 9.2.3; so it is worth checking performance after upgrading to that maintenance release. A proper fix, which has been confirmed to fix the problem the reverted patch fixed without causing a regression, will be included in 9.3.0.