Re: the RAID question, again

From: "scott(dot)marlowe" <scott(dot)marlowe(at)ihs(dot)com>
To: Vincent van Leeuwen <pgsql(dot)spam(at)vinz(dot)nl>
Cc: <pgsql-performance(at)postgresql(dot)org>
Subject: Re: the RAID question, again
Date: 2003-04-22 17:29:07
Message-ID: Pine.LNX.4.33.0304221120490.9550-100000@css120.ihs.com
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-performance

On Tue, 22 Apr 2003, Vincent van Leeuwen wrote:

> On 2003-04-16 19:32:54 -0700, Nikolaus Dilger wrote:
> > One improvement area may be to put all 6 disks into a
> > RAID 10 group. That way you have more I/O bandwith.
>
> A concern I have about that setup is that a large WAL write will have to wait
> for 6 spindles to write the data before returning instead of 2 spindles. But
> as you say it does create way more I/O bandwidth. I think I'll just test that
> when the box is here instead of speculating further :)

Not in a RAID 10. Assuming the setup is:

RAID0-0: disk0, disk1, disk2
RAID0-1: disk3, disk4, disk5
RAID1-0: RAID0-0, RAID0-1

Then a write would only have to wait on two disks. Assuming the physical
setup is one SCSI channel for RAID0-0 and one for RAID0-1, then both
drives can write at the same time and your write performance is virtually
identical to a single drive.

> On 2003-04-16 20:20:50 -0700, Josh Berkus wrote:
> > Nickolaus has a good point. With a high-end Linux server, and a medium-end
> > RAID card, it's sometimes faster to use Linux software RAID than harware
> > raid. Not all the time, though.
>
> I've heard rumors that software raid performs poor when stacking raid layers
> (raid 0 on raid 1). Not sure if that's still true though.

I tested it and was probably the one spreading the rumors. I was testing
on Linux kernels 2.4.9 at the time on a Dual PPro - 200 with 256 Meg RAM
and 6 Ultra Wide 4 gig SCSI drives at 10krpm. I've also tested other
setups.

My experience was that RAID5 and RAID1 were no faster on top of RAID0 then
on bare drives. note that I didn't test for massive parallel
performance, which would probably have better performance with the extra
platters. I was testing something like 4 to 10 simo connects with pgbench
and my own queries, some large, some small.

> My own experiences
> with linux software raid (raid 5 on a low-cost fileserver for personal use)
> are very good (especially in the reliability department, I've recovered from
> two-disk failures due to controllers hanging up with only a few percent data
> loss), although I've never been overly concerned with performance on that
> setup so haven't really tested that.

My experience with Linux RAID is similar to yours. It's always been rock
solid reliable, and acutally seems more intuitive to me now than any of
the hardware RAID cards I've played with. Plus you can FORCE it to do
what you want, whereas many cards refuse to do what you want.

for really fast RAID, look at external RAID enclosures, that take x drives
and make them look like one great big drive. Good speed and easy to
manage, and to Linux it's just a big drive, so you don't need any special
drivers for it.

In response to

Browse pgsql-performance by date

  From Date Subject
Next Message Will LaShell 2003-04-22 17:44:48 Re: the RAID question, again
Previous Message Josh Berkus 2003-04-22 17:18:57 Re: the RAID question, again