User-Agent: Microsoft-Entourage/11.2.5.060620
Date: Sat, 16 Sep 2006 16:46:04 -0700
Subject: Re: RAID 0 not as fast as expected
From: "Luke Lonergan" <llonergan@greenplum.com>
To: "Bucky Jordan" <bjordan@lumeta.com>,
	"Spiegelberg, Greg" <gspiegelberg@cranel.com>,
	"Joshua Drake" <jd@commandprompt.com>
cc: "Craig A. James" <cjames@modgraph-usa.com>,
	pgsql-performance@postgresql.org
Message-ID: <C131DACC.31421%llonergan@greenplum.com>
Thread-Topic: [PERFORM] RAID 0 not as fast as expected
Thread-Index: AcbYe+tsOENR02gVRlOrAK5w1z9kiQASCCsQAAZdXxMABS1nMAA+A5p1
In-Reply-To: <78ED28FACE63744386D68D8A9D1CF5D42099AF@MAIL.corp.lumeta.com>
MIME-Version: 1.0
Content-Type: text/plain;
 charset=us-ascii
Content-Transfer-Encoding: 7bit

Bucky,

On 9/15/06 11:28 AM, "Bucky Jordan" <bjordan@lumeta.com> wrote:

> What other file systems have you had good success with? Solaris would be
> nice, but it looks like I'm stuck running on FreeBSD (6.1, amd64) so
> UFS2 would be the default. Not sure about XFS on BSD, and I'm not sure
> at the moment that ext2/3 provide enough benefit over UFS to spend much
> time on. 

It won't matter much between UFS2 or others until you get past about 350
MB/s.
 
> Also, has anyone had any experience with gmirror (good or bad)? I'm
> thinking of trying to use it to stripe two hardware mirrored sets since
> HW RAID10 wasn't doing as well as I had hoped (Dell Perc5/I controller).
> For a 4 disk RAID 10 (10k rpm SAS/SCSI disks) what would be a good
> target performance number? Right now, dd shows 224 MB/s.

Each disk should sustain somewhere between 60-80 MB/s (see
http://www.storagereview.com/ for a profile of your disk).

Your dd test sounds suspiciously too fast unless you were running two
simultaneous dd processes.  Did you read from a file that was at least twice
the size of RAM?

A single dd stream would run between 120 and 160 MB/s on a RAID10, two
streams would be between 240 and 320 MB/s.
 
> And lastly, for a more OLAP style database, would I be correct in
> assuming that sequential access speed would be more important than is
> normally the case? (I have a relatively small number of connections, but
> each running on pretty large data sets).

Yes.  What's pretty large?  We've had to redefine large recently, now we're
talking about systems with between 100TB and 1,000TB.

- Luke