Message-ID: <4694F8C5.2040005@p-dw.com>
Date: Wed, 11 Jul 2007 17:35:33 +0200
From: Patric de Waha <lists@p-dw.com>
User-Agent: Thunderbird 1.5.0.12 (Windows/20070509)
MIME-Version: 1.0
To: pgsql-performance@postgresql.org
Subject: Two questions.. shared_buffers and long reader issue
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

Hi,
   I've two questions for which I not really found answers in the web.

   Intro:
   I've a Website with some traffic.
   2 Million queries a day, during daylight.
   Postgres is running on a dedicated server  P4 DualCore, 4 Gig Ram.
   Mainly updates on 1 tuple. And more or less complex SELECT statements.
    I noticed that the overall performance of postgres is decreasing 
when one or more long
   readers are present. Where a long reader here is already a Select 
count(*) from table.

   As postgres gets slower an slower, and users still hammering on the 
reload button to get their
   page loaded. Postgres begins to reach max connections, and web site 
is stuck.
   It's not because of a bad schema or bad select statements. As I said, 
a select count(*) on big table is already
   triggering this behaviour.

   Why do long readers influence the rest of the transactions in such a 
heavy way?
   Any configuration changes which can help here?
   Is it a disc-IO bottleneck thing?
 
   Second question. What is the right choice for the shared_buffers size?
   On a dedicated postgres server with 4 Giga RAM. Is there any rule of 
thumb?
   Actually I set it to +-256M.

 
thanks for any suggestions.

Patric


My Setup:

Debian Etch
PSQL: 8.1.4

WAL files are located on another disc than the dbase itself.

max_connections = 190
shared_buffers = 30000            
temp_buffers = 3000               
work_mem = 4096                
maintenance_work_mem = 16384          
fsync = on                           
wal_buffers = 16                 
effective_cache_size = 5000