IDE channels not independent in PIIX4E?

IDE channels not independent in PIIX4E?

[GrantC]

Greetings,

I have an old but reliable Intel SE440BX-2 with 440BX and PIIX4E 
chipset.  120 + 80 GB hard drives on each channel as master, very 
disappointing md software RAID0 or RAID1 performance.

Slackware 10.1 with 2.4.29 or 2.4.30pre2, testing back to 2.4.21.

One thing I noticed is that if I start a 

  dd if=/dev/zero bs=4k of=/dev/hdX

on each drive, then Ctrl-C stop each, one terminal will wait for 
the other drive to complete the write before starting its turn 
to complete the pending write.

It seems that somewhere what I expected to be two parallel data 
streams are serialised.  As the chipset only does udma33 I was 
hoping to get an OS performance gain with RAID0.  Box has 512MB
memory and rarely touches swap.  

Is this a limit of the hardware or did I miss some kernel tuning?

Have benchmarked with tiobench -- that tells me chunk-size has 
little effect, software RAID costs about 8% CPU, and sequential 
writes cost roughly 20%, 40% and 60% for ext2, ext3 and reiserfs.

Filesystem sequential read speed improved from 17 to 25 MB/s for 
RAID0, but no improvement in write speed.

Raw (non-filesystem) dd speeds are about 27MB/s read, 20MB/s write
to non-RAID partition.  All testing done on same partitions, so 
no confusion with fast or slow end of drive.  The drives are 
Seagate 7200.7 series with 2MB cache.  

Any pointers on tuning options appreciated, the Intel datasheets 
don't really tell me much.  As for reading the kernel source, I 
have no idea where to start, not much under Docs, and I dunno 
which or which way to turn the 'knobs'.

Even search terms would be welcome :)

Thanks,
Grant.

Re: IDE channels not independent in PIIX4E?

[Greg Freemyer]

I can't speak to your specific motherboard/chipset, but the
performance testing I have done shows that the onboard IDE controllers
do not have two independent channels, so using both channels does not
provide twice the speed of a single channel, not even close.

True under both Linux and Windows2K.

OTOH, I have tested 3 or 4 PCI-based Dual Channel IDE controllers and
they work much better from a performance perspective.  I recently
tested both the SIIG dual-channel card and Adaptec dual-channel card. 
They were both in the $50-$60 range and neither has given me any
problems, but I have heard people complain about the SIIG.

HTH
Greg


On Fri, 25 Feb 2005 16:47:46 +1100, GrantC <grant_nospam@dodo.com.au> wrote:
> Greetings,
> 
> I have an old but reliable Intel SE440BX-2 with 440BX and PIIX4E
> chipset.  120 + 80 GB hard drives on each channel as master, very
> disappointing md software RAID0 or RAID1 performance.
> 
> Slackware 10.1 with 2.4.29 or 2.4.30pre2, testing back to 2.4.21.
> 
> One thing I noticed is that if I start a
> 
>   dd if=/dev/zero bs=4k of=/dev/hdX
> 
> on each drive, then Ctrl-C stop each, one terminal will wait for
> the other drive to complete the write before starting its turn
> to complete the pending write.
> 
> It seems that somewhere what I expected to be two parallel data
> streams are serialised.  As the chipset only does udma33 I was
> hoping to get an OS performance gain with RAID0.  Box has 512MB
> memory and rarely touches swap.
> 
> Is this a limit of the hardware or did I miss some kernel tuning?
> 
> Have benchmarked with tiobench -- that tells me chunk-size has
> little effect, software RAID costs about 8% CPU, and sequential
> writes cost roughly 20%, 40% and 60% for ext2, ext3 and reiserfs.
> 
> Filesystem sequential read speed improved from 17 to 25 MB/s for
> RAID0, but no improvement in write speed.
> 
> Raw (non-filesystem) dd speeds are about 27MB/s read, 20MB/s write
> to non-RAID partition.  All testing done on same partitions, so
> no confusion with fast or slow end of drive.  The drives are
> Seagate 7200.7 series with 2MB cache.
> 
> Any pointers on tuning options appreciated, the Intel datasheets
> don't really tell me much.  As for reading the kernel source, I
> have no idea where to start, not much under Docs, and I dunno
> which or which way to turn the 'knobs'.
> 
> Even search terms would be welcome :)
> 
> Thanks,
> Grant.
> 
> -
> To unsubscribe from this list: send the line "unsubscribe linux-ide" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>