RAID 1 vs 5
Edward Arenberg
(Compaq 4214). The array has 6 x 18.4 GB 10K Ultra3 drives with an Ultra2
LVD single-channel I/O card. The E450 does not have a SCSI card capable of
interfacing to the array yet, so I am looking at several options and was
hoping to get some advice.
The E450 will be running both a database and several Web servers, as well as
some miscellaneous tasks. The plan is to split the drive array for both the
database and Web pages (HTML, JS, and images - yes there will be a *lot* of
Web data). The database will be in a raw partition. I hope that's enough
background information.
I have been considering the merits of setting up the array as either
a) 3 x 18 GB mirrored (RAID 1) disks, or
b) a large RAID 5 configuration with 5 drives (72 GB + checksums) and 1 hot
spare.
I am also considering either a host (software) controller, or an embedded
(hardware / PCI bus) controller.
Any advice on this would be greatly appreciated. Thank you in advance.
Ed
Akop Pogosian
> I have an E450 quad-400 box I'm setting up with an external disk array
> (Compaq 4214). The array has 6 x 18.4 GB 10K Ultra3 drives with an Ultra2
> LVD single-channel I/O card. The E450 does not have a SCSI card capable of
> interfacing to the array yet, so I am looking at several options and was
> hoping to get some advice.
You can probably connect it using using Sun's single-ended PCI
UltraSCSI card or using the SCSI port on your E450 if it is not in
use. Those support only 40MB/second mode. I am not aware of Sun cards
that support LVD SCSI, but you can get one from LSI
(www.intraserver.com is their online store).
> The E450 will be running both a database and several Web servers, as well as
> some miscellaneous tasks. The plan is to split the drive array for both the
> database and Web pages (HTML, JS, and images - yes there will be a *lot* of
> Web data). The database will be in a raw partition. I hope that's enough
> background information.
> I have been considering the merits of setting up the array as either
> a) 3 x 18 GB mirrored (RAID 1) disks, or
> b) a large RAID 5 configuration with 5 drives (72 GB + checksums) and 1 hot
> spare.
If you're not using a decent hardware RAID controller the b) option
should be out of question if you need decent write performance (e.g
faster than 2-3MB/second for sequential writes or well below than
1MB/second for random writes) Moreover, since you chose to use a hot
spare disk, the b) option gives only 18GB of extra disk space at a
huge performance cost. So, it is not worth it IMO.
> I am also considering either a host (software) controller, or an embedded
> (hardware / PCI bus) controller.
> Any advice on this would be greatly appreciated. Thank you in advance.
A decent hardware RAID controller can make a big difference, specially
for RAID5 writes. Consider getting a hardware RAID controller if you
need to use RAID5. Obviously, you need a controller that is supported
on solaris/sparc, most of PCI RAID controllers on the market are not
supported on Solaris/sparc. Sun sells one but I am not sure if they
support it with anything other than Sun's own disks and disk array
(though it could work with your array anyways)
The disksuite is Sun's software RAID implementation and it works
fairly well. Since this is a software implementation, RAID5 write
performance noticeably slower than striping and mirroring. However,
the read performance is supposed to be close to what you get with
plain disk striping. Disksuite reference manual suggests to consider
using RAID5 when reads are the dominant I/O operation on the file
system. I would certainly consider using software RAID5 for read
intensive file systems (e.g. web pages, images, etc) which are not
likely to be impacted by slow write performance.
--
Akop Pogosian
This space has been intentionally left blank.
Robin Garner
news:9rg7jc$2ala$1...@agate.berkeley.edu:
> Edward Arenberg <aren...@ep.com> wrote:
>> The E450 will be running both a database and several Web servers, as
>> well as some miscellaneous tasks. The plan is to split the drive
>> array for both the database and Web pages (HTML, JS, and images - yes
>> there will be a *lot* of Web data). The database will be in a raw
>> partition. I hope that's enough background information.
>
>> I have been considering the merits of setting up the array as either
>> a) 3 x 18 GB mirrored (RAID 1) disks, or
>> b) a large RAID 5 configuration with 5 drives (72 GB + checksums) and
>> 1 hot spare.
I'd think carefully about the hot spare - I recently did the availability
calculations, and believe that for a RAID-5 set with less than 15-20 disks,
the hot spare is overkill.
Using the very conservative assumptions that the MTBF of a disk drive is
200,000 hours, it takes 2 hours to merge the hot spare into the set and 24
hours to repair a failed disk, the availability percentages are:
5-disk RAID with hot spare: 99.9999976%
6-disk RAID with no hot spare: 99.999957%
and the mean time between failures is:
5-disk RAID with hot spare: 114,077 years
6-disk RAID without hot spare: 6,338 years
Given that the MTBF of the CPU, RAID controller, cables etc is probably
much less than 6,000 years, I think you'll find the extra 18GB is more
worthwhile
Edward Arenberg
news:9rg7jc$2ala$1...@agate.berkeley.edu...
> Edward Arenberg <aren...@ep.com> wrote:
> > I have an E450 quad-400 box I'm setting up with an external disk array
> > (Compaq 4214). The array has 6 x 18.4 GB 10K Ultra3 drives with an
Ultra2
> > LVD single-channel I/O card. The E450 does not have a SCSI card capable
of
> > interfacing to the array yet, so I am looking at several options and was
> > hoping to get some advice.
>
> You can probably connect it using using Sun's single-ended PCI
> UltraSCSI card or using the SCSI port on your E450 if it is not in
> use. Those support only 40MB/second mode. I am not aware of Sun cards
> that support LVD SCSI, but you can get one from LSI
> (www.intraserver.com is their online store).
>
I found a few LVD SCSI cards - LSI, as you mentioned, Antares, and Partners
(partnersdata.com) have cards that work on a Sun. I figured that the SE
SCSI on the Sun was HVD and would not work with the LVD interface on the
array. If that's not correct, please let me know.
>
> A decent hardware RAID controller can make a big difference, specially
> for RAID5 writes. Consider getting a hardware RAID controller if you
> need to use RAID5. Obviously, you need a controller that is supported
> on solaris/sparc, most of PCI RAID controllers on the market are not
> supported on Solaris/sparc. Sun sells one but I am not sure if they
> support it with anything other than Sun's own disks and disk array
> (though it could work with your array anyways)
>
I would definitely go with a hardware RAID controller if I decided to go
with the RAID 5 configuration. The Sun card seems way overpriced. The only
other I have found is from Infortrend (IFT-2101U2A/B - single/dual channel).
However I am concerned with how well this will really work - I have heard
there are hardware RAID issues with Solaris 8 that I don't think I want to
deal with.
Chris Newport
The built-in interface is single ended Ultra Wide SCSI, which should
work
OK with LVD devices as these should have a compatibility mode.
DO NOT use Sun's Differential SCSI cards, they are HVD and WILL damage
your equipment.
If you need an extra card I would go for the Antares.
Edward Arenberg
news:3BDDAAA9...@NOSPAM.netunix.com...
Thanks for the advice. I'm aware of the HVD - LVD issue. The E450 also has
a Sun differential card, but that will be used to connect to a tape array,
which has an HVD interface.
Any reason for the Antares over the LSI? It looks like the LSI parts are
made by IntraServer Technologies.
Chris Newport
> Any reason for the Antares over the LSI? It looks like the LSI parts are
> made by IntraServer Technologies.
Antares make most of Sun's PCI cards, so the drivers are already there.
Solaris sees the Antares dual LVD card as a Sun dual fast scsi.
Spam Collector
only about 365.25x24= 8766 hours in a year, that yields an average
MTBF of about 22.8 years. Since the M stands for 'mean', for every
drive that lasts 40 years (fat chance) there is one that dies almost
immediately. From my experience, you are lucky if a drive lasts 10
years, and in a 10 drive shelf you will probably lose at least one
in the first year or two.
The hot spare is for betting that a second drive will fail before
you have a chance to replace the first one and finish a rebuild.
If the machine is near your desk and you have spare drives laying
around it is probably a waste. If it is at a colo facility across
the country and a replacement drive would have to ordered if a
failure occurs, then a hot spare may be well worthwhile if uptime
is of any importance.
Frank
On Mon, 29 Oct 2001 05:50:35 GMT, Robin Garner
<robin....@delete.iname.com> wrote:
>Akop Pogosian <akopps...@ocf.berkeley.edu> wrote in
>news:9rg7jc$2ala$1...@agate.berkeley.edu:
>
Robin Garner
news:slrn9tuhsr.63t...@sphinx.artair.com:
> Take those MTBF caculations with a big grain of salt.
Are you querying my arithmetic, or the usefulness of the MTBF figure ?
> Since there's
> only about 365.25x24= 8766 hours in a year, that yields an average
> MTBF of about 22.8 years. Since the M stands for 'mean', for every
> drive that lasts 40 years (fat chance) there is one that dies almost
> immediately. From my experience, you are lucky if a drive lasts 10
> years, and in a 10 drive shelf you will probably lose at least one
> in the first year or two.
Umm ... if the MTBF of a disk is 10 years, and you have 10 disks, then your
expectation is that one disk will fail every year. This is basic
probability. Published availability statistics are generally based on
observing the failure rate of a large sample of disks over a (relatively)
short time period.
> The hot spare is for betting that a second drive will fail before
> you have a chance to replace the first one and finish a rebuild.
> If the machine is near your desk and you have spare drives laying
> around it is probably a waste. If it is at a colo facility across
> the country and a replacement drive would have to ordered if a
> failure occurs, then a hot spare may be well worthwhile if uptime
> is of any importance.
My calculation is based on the assumption that the MTTR (mean time to
repair) of a failed disk is 24 hours (I assume that anyone interested in
availability will have a maintenance contract), and that a hot-spare takes
2 hours to sync. with the rest of the RAID set, in effect lowering the MTTR
from 24 hours to 2 hours.
My observation is that vendors will sell you the hot spare based on FUD and
arm-waving, whereas a quantitative analysis will show it up as a waste of
money.
Regards,
Robin