software raid issues -- possible kernel I/O problem?

software raid issues -- possible kernel I/O problem?

[Chris Friesen]

I've just installed some new ata100 hard drives and matching controller
card in a machine running 2.4.16, and have been experimenting with
software raid.  I've run into some interesting issues with regards to
transfer speeds.

I've got md0 as a raid0, and md1 as a raid1.  hda is an older ata33
drive connected as master to the motherboard, while hde and hdg are the
new ata100 drives each connected as master on one of the two channels on
the adapter (but the adapter only has one IRQ, so I don't know how that
will affect things...).

Using hdparm, I get the following results:

md0: 98.46/42.67
md1: 98.46/29.22
hda: 98.46/9.1
hde: 97.71/31.37
hdg: 96.24/30.92

I also tried some simultaneous runs, with results as follows:
hde and hdg:  50.20/23.27  and  53.56/21.77
hde and hda:  50.59/29.09  and  55.90/9.17


So, my observations are as follows:

1) It seems as though I can't get aggregate burst speeds up above about
100MB/s no matter what I do, even when it's on separate interfaces with
separate IRQs.  Is this running into the limitation of the PCI bus?  I'm
also somewhat confused as to how my old ata33 drive managed to score
nearly 100MB/s burst speed, as well as how some people are claiming
scores of 160MB/s on a ata100 drive (and why I'm not getting that on
mine).

2) Similarly, actual read speads appear limited to an agregate of about
45MB/s in both the raid-0 and simultaneous runs. Why am I not getting
twice the throughput of the single drive case?  Could this be due to the
ata100 controller only using a single IRQ?

3) It doesn't appear as though raid-1 reads are being parallelized. 
This surprised me, as I thought that raid-1 was supposed to come close
to raid-0 in terms of read performance.  Anyone have any ideas about why
this isn't happening?

I'd appreciate any comments you have, or if this isn't the right place
to talk about this, then a redirection to the appropriate forum.

Thanks,

Chris

Re: software raid issues -- possible kernel I/O problem?

[Chris Friesen]

A number of people have privately pointed out that hdparm -T doesn't
actually go to the disk at all.  Guess I should RTFM...I though that
this was reading from the disk's cache, not linux's cache.  Oops.

I'm still kind of curious why raid-1 reads don't seem to get any
performance increase over reads from a single disk.  Any ideas?

Chris

Re: software raid issues -- possible kernel I/O problem?

[Jakob Østergaard]

On Fri, Dec 07, 2001 at 05:39:53PM -0500, Chris Friesen wrote:
> 
> A number of people have privately pointed out that hdparm -T doesn't
> actually go to the disk at all.  Guess I should RTFM...I though that
> this was reading from the disk's cache, not linux's cache.  Oops.
> 
> I'm still kind of curious why raid-1 reads don't seem to get any
> performance increase over reads from a single disk.  Any ideas?

For one single large sequential read, the current RAID-1 code will
not show any significant benefit over the single-disk case.

However, if you read many smaller files, or have multiple concurrent
readers, you should see a good speedup.

Try running two bonnies at the same time.

-- 
................................................................
:   jakob@unthought.net   : And I see the elder races,         :
:.........................: putrid forms of man                :
:   Jakob Østergaard      : See him rise and claim the earth,  :
:        OZ9ABN           : his downfall is at hand.           :
:.........................:............{Konkhra}...............:

Re: software raid issues -- possible kernel I/O problem?

[Roy Sigurd Karlsbakk]

> <snip />
> So, my observations are as follows:
>
> 1) It seems as though I can't get aggregate burst speeds up above about
> 100MB/s no matter what I do, even when it's on separate interfaces with
> separate IRQs.  Is this running into the limitation of the PCI bus?  I'm
> also somewhat confused as to how my old ata33 drive managed to score
> nearly 100MB/s burst speed, as well as how some people are claiming
> scores of 160MB/s on a ata100 drive (and why I'm not getting that on
> mine).

The 100MB/s is probably the PCI bus, yes. The theoretical limit of a
32bit,33MHz PCI bus is 33^6*32/8 = 132.000.000 = ~ 125MB/s. In practice,
there's always some overhead.

An ATA33 drive can't go over 33MB/s, as the bus don't deliver any more.
an ATA100 bus can give you 100MB/s, but there is currently no drive
available that can give you that. The fastest drive I've read about, is
the Western Digital WD1200BB (120 gig). They brag about pushing 525Mbit/s
from buffer to disk. That's some 65 megs per sec, which is pretty amazing.

roy

--
Roy Sigurd Karlsbakk, MCSE, MCNE, CLS, LCA

Computers are like air conditioners.
They stop working when you open Windows.