Re: Performance of Seq Scan from buffer cache

[Matt Daw <matt@xxxxxxxxxxxxxxxxxxx>]

Ugh, never mind. I ran ltrace and it's spending 99% of its time in gettimeofday.

select count(*) from notes;

  count  

---------

 1926207

(1 row)

Time: 213.950 ms

explain analyze select count(*) from notes;

                                                       QUERY PLAN                                                       

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

 Aggregate  (cost=99274.59..99274.60 rows=1 width=0) (actual time=2889.325..2889.325 rows=1 loops=1)

   ->  Seq Scan on notes  (cost=0.00..94459.07 rows=1926207 width=0) (actual time=0.005..1475.218 rows=1926207 loops=1)

 Total runtime: 2889.360 ms

(3 rows)

Time: 2889.842 ms

On Tuesday, 21 August, 2012 at 3:57 PM, Matt Daw wrote:

Howdy. I'm curious what besides raw hardware speed determines the performance of a Seq Scan that comes entirely out of shared buffers… I ran the following on the client's server I'm profiling, which is otherwise idle:

EXPLAIN (ANALYZE ON, BUFFERS ON) SELECT * FROM notes;

Seq Scan on notes  (cost=0.00..94004.88 rows=1926188 width=862) (actual time=0.009..1673.702 rows=1926207 loops=1)

   Buffers: shared hit=74743

 Total runtime: 3110.442 ms

(3 rows)

… and that's about 9x slower than what I get on my laptop with the same data. I ran stream-scaling on the machine and the results seem reasonable (8644.1985 MB/s with 1 core -> 25017 MB/s with 12 cores). The box is running 2.6.26.6-49 and postgresql 9.0.6.

I'm stumped as to why it's so much slower, any ideas on what might explain it… or other benchmarks I could run to try to narrow down the cause?

Thanks!

Matt