Re: 5.3-RC1 poor ATA perfomance

> Date: Wed, 27 Oct 2004 15:25:31 +0200
> From: Willem Jan Withagen <wjw_at_withagen.nl>
> Sender: owner-freebsd-current_at_freebsd.org
> 
> Alexander S. Usov wrote:
> 
> > Hi!
> > 
> > I have hit a very strange problem with a poor ata perfomance.
> > 
> > First some numbers:
> > === FreeBSD 5.3-RC1
> > # dd if=/dev/ad0 of=/dev/null bs=1m count=512
> > 536870912 bytes transferred in 23.891834 secs (22470896 bytes/sec)
> > # dd if=/dev/zero of=/dev/ad0s1 bs=1m count=512
> > 536870912 bytes transferred in 35.413467 secs (15160078 bytes/sec)
> > 
> > === Linux 2.6.8-1 (OpenBSD also gives similar numbers)
> > # dd if=/dev/hda of=/dev/null bs=1M count=512
> > 536870912 bytes transferred in 22.272891 seconds (24104231 bytes/sec)
> > # dd if=/dev/zero of=/dev/hda1 bs=1M count=512
> > 536870912 bytes transferred in 23.574157 seconds (22773706 bytes/sec)
> > 
> > As you see the writing is quite slower.
> > 
> > Some details on hardware:
> > atapci0: <Intel ICH4 UDMA100 controller> port 0x1860-0x186f,0x376,
> > 0x170-0x177,0x3f6,0x1f0-0x1f7 at device 31.1 on pci0
> > ata0: channel #0 on atapci0
> > ad0: 38154MB <TOSHIBA MK4021GAS/GA224A> [77520/16/63] at ata0-master UDMA100
> > 
> > ad1 is absent.
> > 
> 
> Arrrggghhhh, please not all over again.
> 
> We've just sort of terminated the previous thread about this:
>    http://lists.freebsd.org/pipermail/freebsd-current/2004-October/040474.html
> And it is a LONG thread.
> 
> One of the more essentail remarks made bij PHK was that this type of testing 
> is close to no use/good. And to be honest: previously discussed number were 
> even worse. And were probably dues to hardware which was designed/backed on 
> monday-morning.
> First problem is that you need to determine what you want to do with your system.
> 
> Please read the thread.

While reading the thread is a good idea (though it does digress into a
very interesting tutorial on RAID), the problem covered in that stream
is NOT the problem seen here.

I, too, have reported the loss of write speed in V5 vs. V4.

Yes, dd(1) is really only a good test of dd(1), but I have real need to
do dd(1) a lot and the slow-down is most pronounced.

Søren has not been able to reproduce this, but I see it on my ICH3 laptop
every morning when I copy my hard drive out. This problem ONLY impacts
writes, not reads. It is only present on V5 and I think it appeared
with ATAng. It costs me about 25 minutes a day, so it IS significant to
me.

I have not run bonnie++ yet, but will try to do so today.

For the record, IBM T30 with Hitachi (nee IBM) 5400 RPM 40 GB disks with
dd operation run in single user mode. One disk per IDE channel, bot
master. 
-- 
R. Kevin Oberman, Network Engineer
Energy Sciences Network (ESnet)
Ernest O. Lawrence Berkeley National Laboratory (Berkeley Lab)
E-mail: oberman_at_es.net			Phone: +1 510 486-8634