Message-ID: <4BEEFA75.6090708@agliodbs.com>
Date: Sat, 15 May 2010 12:48:05 -0700
From: Josh Berkus <josh@agliodbs.com>
User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US;
	rv:1.9.1.9pre) Gecko/20100217 Shredder/3.0.3pre
MIME-Version: 1.0
To: pgsql-performance@postgresql.org
Subject: Re: Benchmark with FreeBSD 8.0 and pgbench
References: <28569544.post@talk.nabble.com>
In-Reply-To: <28569544.post@talk.nabble.com>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit

Joao,

Wow, thanks for doing this!

In general, your tests seem to show that there isn't a substantial 
penalty for using ZFS as of version 8.0.

If you have time for more tests, I'd like to ask you for a few more tweaks:

(1) change the following settings according to conventional wisdom:
	wal_buffers = 8MB
	effective_cache_size = 9GB
	checkpoint_segments = 32
	on ZFS only: full_page_writes=off

(2) What scale were you using for the pgbench database?  I didn't see it 
in the e-mail.  It would be worth testing:
	s = 10 (small database, in memory)
	s = 500 (7GB, ram mostly full)
	s = 1000 (14GB, slightly larger than ram)
	s = 3000 (43GB, much larger than ram)

If you were only testing a small size in your runs, then the only 
Filesystem behavoir you were testing was the transaction log.

(3) Try a ZFS 128K record size

(4) Centos/Ext3 appears to have had better staying power with high 
numbers of clients.  Can you continue testing with 50, 100 and 200 
clients on that combination?  And with data=writeback,noatime on Ext3?

-- 
                                   -- Josh Berkus
                                      PostgreSQL Experts Inc.
                                      http://www.pgexperts.com