To: "Jean-David Beyer" <jeandavid8@verizon.net>
Cc: <pgsql-performance@postgresql.org>
Subject: Re: random_page_costs - are defaults of 4.0 realistic for SCSI RAID 1
In-Reply-To: <46E676D6.70807@verizon.net> (Jean-David Beyer's message of
 "Tue\,
	11 Sep 2007 07\:07\:02 -0400")
User-Agent: Gnus/5.11 (Gnus v5.11) Emacs/22.1 (gnu/linux)
References: <C30B1716.4236C%llonergan@greenplum.com>
	<87lkbdlkxw.fsf@oxford.xeocode.com> <46E676D6.70807@verizon.net>
From: Gregory Stark <stark@enterprisedb.com>
Organization: EnterpriseDB
Date: Tue, 11 Sep 2007 19:16:03 +0100
Message-ID: <87zlzt9ifw.fsf@oxford.xeocode.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii

"Jean-David Beyer" <jeandavid8@verizon.net> writes:

> Gregory Stark wrote (in part):
>
>> The extra spindles speed up sequential i/o too so the ratio between sequential
>> and random with prefetch would still be about 4.0. But the ratio between
>> sequential and random without prefetch would be even higher.
>> 
> I never figured out how extra spindles help sequential I-O because
> consecutive logical blocks are not necessarily written consecutively in a
> Linux or UNIX file system. They try to group a bunch (8 512-bit?) of blocks
> together, but that is about it. So even if you are reading sequentially, the
> head actuator may be seeking around anyway. 

That's somewhat true but good filesystems group a whole lot more than 8 blocks
together. You can do benchmarks with dd and compare the speed of reading from
a file with the speed of reading from the raw device. On typical consumer
drives these days you'll get 50-60MB/s raw and I would expect not a whole lot
less than that with a large ext2 file, at least if it's created all in one
chunk on a not overly-full filesystem. 

(Those assumptions is not necessarily valid for Postgres which is another
topic, but one that requires some empirical numbers before diving into.)

-- 
  Gregory Stark
  EnterpriseDB          http://www.enterprisedb.com