Skip to contentSkip to conversation

harddisk speed: 2.4.24 20+% faster then 2.6.3

DISCUSSION Danny ter Haar 2 participants 3 messages Feb 27, 2004

lore
Share & tools
lore
#0 lore raw

I've been checking, rechecking but cannot explain why a "vannilla" debian install kernel (sid-testing install cdrom) is faster then "Feisty Dunnart" on the same hardware.

hardware: Mini-itx with 1Ghz via-C3
200GB udma harddisk.

2.4.24:
Timing buffered disk reads: 150 MB in 3.01 seconds = 49.83 MB/sec

2.6.3 (vanilla):
Timing buffered disk reads: 102 MB in 3.05 seconds = 33.40 MB/sec

2.6.3-mm4:
Timing buffered disk reads: 108 MB in 3.00 seconds = 35.95 MB/sec

Bonnie tests confirm this!

hdparm & bonnie & kern.log for comparison put at: http://dth.net/c3/

Anyone any idea *why* ?

Mayby i'm overlooking something (it's kind of late on this side of the pond) but still...

Open for feedback.

Danny

Signature · 4 lines
 /"\                        | Dying is to be avoided because
 \ /  ASCII RIBBON CAMPAIGN | it can ruin your whole career 
  X   against HTML MAIL     | 
 / \  and POSTINGS          | - Bob Hope
#1 lore raw
Quoted 7 lines · Danny ter Haar <[email protected]>, on Fri Feb 27, 2004 [12:04:29 AM]…

I've been checking, rechecking but cannot explain why a "vannilla" debian install kernel (sid-testing install cdrom) is faster then "Feisty Dunnart" on the same hardware.

hardware: Mini-itx with 1Ghz via-C3
200GB udma harddisk.

Quoted 9 lines

2.4.24:
Timing buffered disk reads: 150 MB in 3.01 seconds = 49.83 MB/sec

2.6.3 (vanilla):
Timing buffered disk reads: 102 MB in 3.05 seconds = 33.40 MB/sec

2.6.3-mm4:
Timing buffered disk reads: 108 MB in 3.00 seconds = 35.95 MB/sec

	Hi;

	
	By default, my hdparm numbers and my kernel build time

benchmark were worse on all 2.6 vs. 2.4 kernels, until I tried:

/sbin/hdparm -a 4096 /dev/hdx

eg. (typical averages)
2.4.21:
Timing buffered disk reads: 80 MB in 3.07 seconds = 26.06 MB/sec
2.6.0-test5:
Timing buffered disk reads: 68 MB in 3.06 seconds = 22.21 MB/sec
2.6.1: [readahead 4096]
Timing buffered disk reads: 90 MB in 3.02 seconds = 29.82 MB/sec

2.6.0-test10 building 2.4.18-packet (personal config) gcc3.2.3 glibc2.3.2 160g Maxtor, ext3
(2xpII 400)
time (make -j4 clean && make -j4 dep && make -j4 bzImage) [with default read-ahead]

real    9m1.496s
user    16m1.977s
sys     0m45.347s

[read-ahead = 4096]

real    7m2.923s
user    12m6.889s
sys     0m46.190s
	pIIx4, ata33, 160g maxtor

Paul
[email protected]

#2 lore raw
Quoted 10 lines · Paul

By default, my hdparm numbers and my kernel build time benchmark were worse on all 2.6 vs. 2.4 kernels, until I tried:
/sbin/hdparm -a 4096 /dev/hdx
eg. (typical averages)
2.4.21:
Timing buffered disk reads: 80 MB in 3.07 seconds = 26.06 MB/sec
2.6.0-test5:
Timing buffered disk reads: 68 MB in 3.06 seconds = 22.21 MB/sec
2.6.1: [readahead 4096]
Timing buffered disk reads: 90 MB in 3.02 seconds = 29.82 MB/sec

Allready thought about that, tried that but no real

difference:belfish:~# uname -a
Linux babelfish 2.6.3-mm4 #3 Fri Feb 27 14:09:38 CET 2004 i686 GNU/Linux
babelfish:~# /sbin/hdparm -a 4096 /dev/hda
/dev/hda:
setting fs readahead to 4096
readahead    = 4096 (on)
babelfish:~# hdparm -t /dev/hda
Timing buffered disk reads:  110 MB in  3.02 seconds =  36.47 MB/sec

I just installed a news amd64 en few 19" servers.
But none of them showed downgrade in performance when i upped the kernel.

Possibilities left (according to me!)

- scheduler (should i try different setup ?)
- irq problems (in combination with acpi?)
i booted with acpi=off and retested but same results.
- driver, specific the via-onboard.

No luck so far!

Danny

Signature · 4 lines
 /"\                        | Dying is to be avoided because
 \ /  ASCII RIBBON CAMPAIGN | it can ruin your whole career 
  X   against HTML MAIL     | 
 / \  and POSTINGS          | - Bob Hope