Skip site navigation (1)Skip section navigation (2)
Date:      Sun, 8 Nov 1998 09:49:16 +1030
From:      Greg Lehey <grog@lemis.com>
To:        Bill Vermillion <bill@bilver.magicnet.net>, freebsd-fs@FreeBSD.ORG
Subject:   Re: RAID1 Software vs Hardware
Message-ID:  <19981108094916.T499@freebie.lemis.com>
In-Reply-To: <199811071331.IAA10243@bilver.magicnet.net>; from Bill Vermillion on Sat, Nov 07, 1998 at 08:31:26AM -0500
References:  <19981107200618.O499@freebie.lemis.com> <199811071331.IAA10243@bilver.magicnet.net>

next in thread | previous in thread | raw e-mail | index | archive | help
On Saturday,  7 November 1998 at  8:31:26 -0500, Bill Vermillion wrote:
> Greg Lehey recently said:
>> On Friday,  6 November 1998 at 19:42:15 -0500, Bill Vermillion wrote:
>>> Christopher Nielsen recently said:
>>>
>>>> On Fri, 6 Nov 1998, Bagnara Stefano wrote:
>>>
>>>>> Date: Fri, 6 Nov 1998 17:35:58 +0100 From: Bagnara Stefano
>>>>> <bago@datasail.it> To: freebsd-fs@FreeBSD.ORG Subject: RAID1
>>>>> Software vs Hardware                                   ^^^^^
>
>>>>> I need to implement a simple raid1 with 2 9gigs SCSI HDD. I    .
>>>>> Ineed a inexpensive solution ... so i was thinking about a     .
>>>>> Isoftware raid1 s it possible? is it too slow? the system will .
>>>>> Ibe only a mail server                                         .
>
>>>> Your really not going to see very good performance with RAID if
>>>> you're using only two spindles (i.e., discs).  ...
>
>>> I've found that I get a 50% throughput increase (typical)when
>>> running RAID 0 with 2 drives.
>
>> That's what theory would tell you.
>
> No - not theory.  Measured in real-life - running HW raid 0 - on a
> clients SCO system.  We needed more speed.  It may be slightly
> under 50% - but it's darn close.

Theory would tell you 50%.  Practice shows it's darn close.  If I had
said "yes, that's what happens", somebody would have looked from the
other perspective and said "not quite".

> I have timed the same drive on SCO and FreeBSD - a 9GB 'cudda, and
> the raw SCO performance through the file-system is in the 3MB/min
> range, while using the FreeBSD file-system - as shipped - no mods,
> etc,. it is between 2 and 3 times faster than SCO's.  

Which operating system?

> Controller is DPT with a hunk of cache.  All timing was done with
> 100MB minimum file sizes usiing iozone.
>
>>> RAID 1 will give no increase in some areas, and writes are slower,
>
>> This depends a lot on the configuration.  You need to write to every
>> copy (plex in vinum jargon), whereas you only need to read from one.
>> In practice, if you're running SCSI disks or IDEs with UDMA on
>> different controllers, you'll delay the writes to the speed of the
>> slowest disk.  On average that's slower than a single disk, but not
>> much.
>
> I've only used HW RAID devices so I can't comment on the other
> implementation.   The RAID Standards Board (I think that's the
> correct name) has permitted a pair of striped drives to be
> designated RAID 0, while in reality it is not a RAID.
>
> They current have 6 classes - 0 thru 5 - and there is a chart
> in Adaptec's book on I/O subsytems listing the pro's'/con's of
> each.  RAID2/3/4 aren't used, and from what I've seen drives that
> use to have spindle sync for byte/sector striping aren't being made
> anymore.  But with drives now at 20MB/sec+ speeds, the old needs
> are gone.

Correct.  You'd be surprised how many products offer RAID-2/3/4,
though.  I think the product manager got a checklist to tick off.  I
have deliberately left these three out of Vinum.

> I can't comment on IDE as I've never used them.

Shouldn't make much difference with only one drive per controller and
UDMA.  Otherwise you'll run into significant contention problems.

>>> but it will boost the read throughput if different files are being
>>> accessed, just as if you load balanced multiple single disks..
>
>> BTW, ccd always reads from the same copy of the data, so this doesn't
>> work.  But in principle you're right.
>
> Reading from two disks for different files is one of the touted
> features of most HW implementations.

Put it this way, I don't know of any other implementation, SW or HW,
which is this primitive.  Vinum has a choice of round robin (default)
or always reading from a specific drive (which can be an advantage if
you have a ramdisk, for example).

Greg
--
See complete headers for address, home page and phone numbers
finger grog@lemis.com for PGP public key

To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-fs" in the body of the message



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?19981108094916.T499>