Kernel 2.6.12 slower than 2.6.11 on Transmeta processor?

folbrich · n00b Joined: 31 Jul 2005 Posts: 9

Hi,
last month I got a new laptop and started with Gentoo and kernel 2.6.11-gentoo-r11. I was quite happy that it booted much faster than Windows. Recently, I upgraded to kernel 2.6.12-gentoo-r6 and noticed that it takes much longer now though, still faster than Windows

. The config is exactly the same for both kernels and I have how to fix this. The new kernel is just slower on my machine (Transmeta Efficeon). I did some tests with the following results:

- Complete startup to KDM, 2.6.11 -> 1:03min, 2.6.12 -> 1:33min
- Kernel Init only , 2.6.11 -> 14 sec, 2.6.12 -> 14sec
- "Calc Module Dependencies", 2.6.11 -> 7sec, 2.6.12 -> 20sec
- Shutdown, 2.6.11 -> 20sec, 2.6.12 -> 33sec
- Compiling a kernel, 2.6.11 -> 17:20min, 2.6.12 -> 22:40min

Does anybody have an idea how I can get back the performance of 2.6.11 with 2.6.12? I tried lots of things but nothing has worked so far. Should I post my kernel config?
Thanks in advance

NeddySeagoon · Posted: Sun Jul 31, 2005 12:39 pm Post subject:

folbrich,

Tell us the steps you went though to migrate from 2.6.11 to 2.6.12.
The 'Exactly' in your post bothers me.
_________________
Regards,

NeddySeagoon

Computer users fall into two groups:-
those that do backups
those that have never had a hard drive fail.

folbrich · n00b Joined: 31 Jul 2005 Posts: 9

Ok, at first I migrated with "make oldconfig" and noticed the slowdown. Today, I took the time to compare both configurations with "make menuconfig". I went through every point in the config and made sure that it's the same. There're only a few additional config options in the 2.6.12 config and I left it untouched. The 2.6.12 kernel has one more module (freq_table.ko) in /lib/modules anyway. I'm not sure if that's related to the problem.

Do you think I should start a new config from scratch?

NeddySeagoon · Posted: Sun Jul 31, 2005 2:03 pm Post subject:

folbrich,

Ah OK. There are too many people who are not aware of the make oldconfig step.
Maybe your CPU is running in a powersave mude under 2.6.12.

Compare the content of /proc/cpuinfo under the two kernels. Are the CPU speeds reported the same?
What about the BOGOMIPS ?
_________________
Regards,

NeddySeagoon

Computer users fall into two groups:-
those that do backups
those that have never had a hard drive fail.

folbrich · n00b Joined: 31 Jul 2005 Posts: 9

Both kernels report the same CPU speed (995MHz) and 1961 bogomips. I also tried to disable Power Management in the 2.6.12 kernel but it didn't help. I don't know where to look now. Maybe it's slow disk IO but hdparm gives me the same results for both kernels.

ycUygB1 · Posted: Sun Jul 31, 2005 5:34 pm Post subject: Same problem

I'm also experiencing the same problem. It is across the board with the ck kernel and the gentoo-r6 kernel.
With the 2.6.11 kernels, my interbench-0.25 results were extraordinary -- only one missed deadline.

But with 2.6.12, the performance is abysmal, some tests miss more than 99% of their deadlines.

Does the gcc compiler version matter much? I'm using gcc-3.3.5 -O2 -march=pentium4.

Harold

folbrich · n00b Joined: 31 Jul 2005 Posts: 9

I think the GCC version doesn't matter. I compiled both kernels with 3.3.5 and the 2.6.11 is still faster. A few minutes ago, I tried to make a new config from scratch and that also didn't change anything.
Maybe I should just stick to 2.6.11 and wait for 2.6.13...

ycUygB1 · Posted: Sun Jul 31, 2005 6:03 pm Post subject: Let me know what you find out.

Well, let me know what you find out. I posted my interbench results to Con Kolivas' mailing list to see what he says.
In the meantime, I'm reading the article Linux: Realtime Benchmarking
on www.kerneltrap.org. Maybe that'll help. Or maybe I'll try installing a precompiled kernel to see if I did something wrong.
Or maybe GCC 4.0

Harold

ycUygB1 · Posted: Sat Aug 06, 2005 1:19 pm Post subject: hdparm

The problem I was having, which is now solved, was that dma was disabled, resulting in very slow disk access times.
I forced dma in the kernel, and then it was fixed.

You can determine if this is the problem by using hdparm /dev/hda. Look at the using_dma field.

At one point I thought it was the power management or timer. So I turned ACPI off and enabled HPET.
These were both mistakes. ACPI is necessary for hyperthreading and no HPET driver exists for my hardware yet.
So now I've got ACPI enabled and the PM timer.

lesshaste · n00b Joined: 04 Aug 2005 Posts: 12

You may be interesed in this from the changelog of 2.6.13-rc5

"commit cdf32eaa4e601b15146e21b6470de00f149ce37f
Author: Eric Lammerts <eric@lammerts.org>
Date: Sun Jul 31 22:34:42 2005 -0700

[PATCH] disable addres space randomization default on transmeta CPUs

We know that the randomisation slows down some workloads on Transmeta CPUs
by quite large amounts. We think it's because the CPU needs to recode the
same x86 instructions when they pop up at a different virtual address after
a fork+exec.

So disable randomization by default on those CPUs.

Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

"

Raphael