Skip to contentSkip to conversation

IDE performance drop between 2.4.23 and 2.6.0

DISCUSSION Jean-Luc Fontaine 2 participants 3 messages Dec 27, 2003

lore
Share & tools
lore
#0 lore raw

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

I solved the problem in a very strange way. Note that the (b) disk performance only improves after readahead has been increased on another
(c) drive! (the (c) drive performance was also increased by to 2.4 levels but is not shown here). I could reliably repeat this behavior after rebooting.

Can any IDE expert explain it?

# hdparm /dev/hdb
~ multcount    =  0 (off)
~ IO_support   =  1 (32-bit)
~ unmaskirq    =  1 (on)
~ using_dma    =  1 (on)
~ keepsettings =  0 (off)
~ readonly     =  0 (off)
~ readahead    = 256 (on)
~ geometry     = 65535/16/63, sectors = 80418240, start = 0
# hdparm -t /dev/hdb
~ Timing buffered disk reads:   34 MB in  3.09 seconds =  11.02 MB/sec
# hdparm -a 4096 /dev/hdb
/dev/hdb: readahead = 4096 (on)
# hdparm -t /dev/hdb
~ Timing buffered disk reads:   34 MB in  3.08 seconds =  11.04 MB/sec
# hdparm -a 4096 /dev/hdc
/dev/hdc: readahead = 4096 (on)
# hdparm -t /dev/hdb
~ Timing buffered disk reads:   46 MB in  3.12 seconds =  14.76 MB/sec

- --
Jean-Luc Fontaine mailto:[email protected] http://jfontain.free.fr/
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.3 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQE/7ZqDkG/MMvcT1qQRAvLKAKC0t3Lsq+B4DBCAwQVadoXVkPxahACfZbfU UBE4kGDSf6WpuMRtsuxSJCU=
=0u8i
-----END PGP SIGNATURE-----

#1 lore raw
Quoted 8 lines · Jean-Luc Fontaine

I solved the problem in a very strange way. Note that the (b) disk performance only improves after readahead has been increased on another
(c) drive! (the (c) drive performance was also increased by to 2.4 levels but is not shown here). I could reliably repeat this behavior after rebooting.

Can any IDE expert explain it?

Looks like a bug. If you don't access hdc, then the read-ahead on hdc shouldn't matter. I wonder if the read-ahead code (either the setting or the reading) gets the value from the wrong queue or something.

Jens? This could certainly explain some of the confusion about performance.. Do you see anything strange?

Linus

Quoted 10 lines
# hdparm -t /dev/hdb
~ Timing buffered disk reads:   34 MB in  3.09 seconds =  11.02 MB/sec
# hdparm -a 4096 /dev/hdb
/dev/hdb: readahead = 4096 (on)
# hdparm -t /dev/hdb
~ Timing buffered disk reads:   34 MB in  3.08 seconds =  11.04 MB/sec
# hdparm -a 4096 /dev/hdc
/dev/hdc: readahead = 4096 (on)
# hdparm -t /dev/hdb
~ Timing buffered disk reads:   46 MB in  3.12 seconds =  14.76 MB/sec
#2 lore raw

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Linus Torvalds wrote:

|
| On Sat, 27 Dec 2003, Jean-Luc Fontaine wrote:
|
|>I solved the problem in a very strange way. Note that the (b) disk
|>performance only improves after readahead has been increased on another
|>(c) drive! (the (c) drive performance was also increased by to 2.4
|>levels but is not shown here). I could reliably repeat this behavior
|>after rebooting.
|>
|>Can any IDE expert explain it?
|
|
| Looks like a bug. If you don't access hdc,

hdc is / for 2.6, whereas hdb is used for 2.4. So hdc was obviously accessed prior to this test.

| then the read-ahead on hdc
| shouldn't matter. I wonder if the read-ahead code (either the setting or
| the reading) gets the value from the wrong queue or something.

Or could this set something in the VIA chipset? I'll take a look in /proc/ide/via and report if needed.

Let me know if you need more tests to be run.

- --
Jean-Luc Fontaine mailto:[email protected] http://jfontain.free.fr/
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.3 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQE/7eOIkG/MMvcT1qQRAjAxAKCr3xDbqELBzY7zsOYCQi/7QsYw7ACeObup auyEo8bDPNHOD79+/I8/xPw=
=rsrt
-----END PGP SIGNATURE-----