SATA drives slow

[Schizoid]

I am using the ATA_PIIX kernel driver for my SATA controller. There are two new, identical 500gb SATA hard disk and a single SATA dvd-rw drive in the system.

This computer is less than two months old and it never seems to perform well when it comes to file operations. For example. I had a backup of my /usr/portage directory which today I finally got around to removing with an rm -r /usr/portage.old command. This is less than 250mb, and it took well over 10 minutes to complete the operation. I washed out my tea pot, boiled some water, and steeped a fresh cup of tea and it still had not finished. I realize there are a zillion small files in there, but come on!

I have been using dd to copy some dvds to one of the partitions, using a simple dd if=/dev/dvd of=image.iso command. dd is kind enough to let me know that this operation is performing at around 11mb/sec. That is around 8x for a dvd drive. Should it not be a lot faster than this? (Using an ext3 partition instead of XFS gets me 12.5mb/sec)

The computer is always very responsive, but waiting over ten minutes to delete 250mb is a joke. There has to be something off here, right?! I thought it might be XFS so I put ext3 on a partition and as you can see above things are not really much better.

[likewhoa]

you might want to consider putting those two drives in a RAID array for performance increases, you might also want to switch to a different iosched and or play with

[Schizoid]

First off I just want to point out that the 12.5mb/sec number was an XFS partition, and the 11mb/sec is for the ext3. I had them mixed up.

I set blockdev --setra 16384 on both disks and the dvdr (all were at 256 previously) and when using the same dd command that gets me 14.7mb/sec on the ext3 (from 11mb/sec). For the XFS partition I get 15.4mb/sec.

Copying the file from the XFS partition to the ext3 (on separate physical disks) I get a rate of about 25.2mb/sec (4465mb in 177 seconds).

These don't seem like the best numbers to me. Do you think the file systems are a factor here? I don't know what else to try to adjust when it comes to SATA.

Thanks for the help.

Oh, and I am using the default cfq scheduler. This system has a quad-core cpu and 2gb of ram. I have never setup a raid array and I don't really want to fiddle with that if I don't have to. If the drives don't seem to be performing well without being in an array the array will just help to mask their poor performance, eh?

[likewhoa]

[Schizoid]

I will give that a shot next time I rebuild the kernel.

Am I crazy or isn't there a way to change scheduler without rebooting?

[gimpel]

[niceness]

[mbar]

[Schizoid]

[Schizoid]

As another example, unpacking source files always takes a very long time. I feel like the computer can compile things faster than it can unpack the source code sometimes.

[Lifthanger]

I have the same problem on my amd64 gentoo. My root is XFS and esspecially untaring sources and deleting them again takes a very long time. Monitoring the process with dstat showed write speeds of under 10MB/s while hdparm -t gives around 95 MB/s.

Maybe we can find a solution together.

[Schizoid]

Ugh, here is the same operation on an ext3 partition in the same system.

[Lifthanger]

hmm seems like an xfs issue..
I'll install another hdd to make some test partitions today. I'll post the results.

[Lifthanger]

the results: untaring the 55MB netbeans sources and rm -rf of the sources afterwards. Kernel is 2.6.23 AMD64 - io-scheduler doesn't make a difference.
The HDD yields 75MB/s in hdparm.

reiserfs3:

real 0m31.201s
user 0m26.055s
sys 0m3.274s

real 0m2.791s
user 0m0.052s
sys 0m2.548s

ext2:

real 0m30.638s
user 0m25.794s
sys 0m2.300s

real 0m0.866s
user 0m0.032s
sys 0m0.416s

xfs:

real 3m37.179s
user 0m26.066s
sys 0m4.524s

real 2m27.426s
user 0m0.052s
sys 0m2.080s


This last test is on a software raid 5 array with 32bit kernel. hdparm nets around 140MB/s but on heavy access it gets cpu restricted, or at least that was my impression.

xfs:

real 1m47.089s
user 0m48.151s
sys 0m7.888s

real 0m21.938s
user 0m0.184s
sys 0m4.876s


Seems to me as if xfs and AMD64 don't like each other very much... I'm open to other conclusions though.

[nbkolchin]

Try this guide for XFS tuning: http://www.everything2.com/index.pl?node_id=1479435