Hi,

I waited 2 hours (during linux-il's prime time) and no useful answer
arrived, so I will share my little experience.

On Thu, Jul 29, 2004 at 09:35:22PM +0300, Eran Tromer wrote:
> Ahoy,
> 
> I'd appreciate advise on the following.
> 
> My main workstation currently has two 80GB ATA drives. One is used
> mostly for backups of the other, so I have just about 80GB of effective
> work space. I need more space, and more speed, so I plan to buy a 160GB
> ATA/133 drive and configure the three drives as follows:
> * 160GB 3-way RAID5 via Linux's software RAID, using the two existing
> 80GB disks and half of the new 160GB disk.
> * The other 80GB of the 160GB disk used as a separate partition for
> less important (i.e., easily recoverable) stuff.
> 
> Is there some problem with using half-a-disk for the RAID array?

Not that I know of.

> How bad
> is it that the three 80GB partitions making up the RAID5 reside on disks
> of different makes and models, contrary to common wisdom?

I do not think it's bad at all.

> Assuming low
> load on the extra 80GB partition, is there some reason to use RAID4
> (with the parity on the 160GB disk) instead of RAID5?

Probably, but I don't know.
The main thing to consider about performance is that modern disks are
faster in their start and slower in their end, with various speeds
(around 10-20 different zones) in between, with the end around half
the speed of the start. Also, newer disks tend to be faster than older
ones, but that's not accurate. In "speed" I refer to throughput only,
not latency (seek time + rotational latency, the latter a direct result
of the RPM). My experience shows (all speeds are in the start) that new
40GB disks are around 45MB/s (old are around 30MB/s), and new 80GB+ ones
are around 55MB/s (I measured several different 80GB, but only few 120
and only one 200GB). The fastest disk I measured until today, BTW, was
a 72GB SCSI disk, 10K RPM (or maybe even 15K, don't remember and don't
have model), which did 69MB/s. Fast. I was pretty disappointed with the
120 and 200, expecting them to be faster than the 80's, but that's life.

This means that the start of your 80's and 160's will be the same speed,
and so there is no big point in giving a different role to one of them,
but the ends of the 80 are going to be much slower than the middle of
the 160 one. So in terms of speed, the best you can do is have more than
one raid array - e.g. slice the 80's to 2 or even 3 partitions, create
same-size partitions on the 160, and raid them. Much less convinient,
though. Speed costs.

I also want to say that I have no idea what speed you are going to get
_in_practice_, and if you are interested and have the time you should
benchmark.

To check disk speed I use zcav (from bonnie). I never benchmarked
seriously a filesystem, but bonnie should do this too.

You can get my quite-large collection of zcav outputs (I know it
sounds pathetic, and it is), at
<http://www.cs.tau.ac.il/~didi/disk-speed.tgz> (a 1.2MB file).
You'll find there, among other things, in the fox/ dir, runs of
zcav on two disks concurrently (in various connections - both PATA,
one through a SATA converter, etc.). I also tried such a thing with
dd bs=1024k count=1000 (no zcav), two PCI IDE RAID and on-board,
results in </~didi/raid.tgz>.

> What's the best
> RAID chunk size for a workload typically consisting of 2-3 heavy batch
> jobs plus normal interactive desktop use (I was thinking of 128KB or
> 256KB)?

I have no idea, I use 4KB, and would love to hear educated opinions.

> Is there a really good reason to get an IDE RAID PCI card?

Not really. There are basically two reasons to do this: One is that
you potentially get more throughput (because connecting 2-3 fast disks
to the same on-board controller will degrade throughput significantly,
by my measurements), the other is os-independent-raid-configuration -
that is, the raid conf is kept on the card, and the driver reads it
from there, and not from /etc or the registry. The last one is an
issue only if you dual-boot. Note (as others said on this list in the
past), that cheap ones are _software_ RAID - the driver does the raid.
Hardware raid starts at around $500 per card.

> How
> much trouble should I expect from mkinitrd, Knoppixen, etc?

I don't know. I know there is support, I know it's not as seamless as
with regular disks, but I did not try it. I only did raid to data
volumes, not "system" ones. But I do not think it's more than a few
hours for an experienced sysadmin.

Good luck, and tell us your results (if/when you'll have some),
-- 
Didi


=================================================================
To unsubscribe, send mail to [EMAIL PROTECTED] with
the word "unsubscribe" in the message body, e.g., run the command
echo unsubscribe | mail [EMAIL PROTECTED]

Reply via email to