[Solved] Kernels 2.6.35+ very slow on an i7.

[drescherjm]

I have been running 2.6.34 for a month or so without issue however 2.6.35 is extreamly slow. Slow to the point that typing in a console, ssh, GUI reminds me of a 300 baud modem and this is a 3.0GHz i7. Normal booting in 2.6.34 or lower is 20 to 30 seconds. I have 300+ second boots on 2.6.36-rc3.

I have copied my kernel config from 2.6.34 to 2.6.35 (and above) just like I have done for the hundreds of other kernel upgrades I have done (I still manage dozens of gentoo boxes home and work). Anyways during the boot I have noticed that all is well up until the ahci driver is loaded. There does not appear to be anything unusual in the dmesg well except for taking much longer.

Here is a part of the init from 2.6.34

[bjlockie]

If you use X, recompile the input drivers.

Maybe gcc debugging got turned on?
_________________
AMD FX6100 CPU, 16 GiB RAM, OCZ Vertex 3 SSD
ASRock 970 Extreme3 motherboard with S/PDIF audio
Galaxy-NVidia GeForce 8800GT video card, Cyber Power CP550HG USB UPS

[drescherjm]

[drescherjm]

With a pappy-seed for 2.6.35-gentoo-r5 I was able to get past the slow typing however there still is a lot of slowness. I mean it still took 170 seconds to boot to the console and it is still more than 30 seconds before I get to the init. Compare that to 20s or so total boot time in 2.6.34 with way more stuff activated..
_________________
John

My gentoo overlay
Instructons for overlay

[Naib]

does this occur with any linux LiveCD's eg Pentoo, SysrescueCD, Ubuntu. ie is this a fundamental issue with i7 at teh moment or is it just a configuration issue

Have you tried disabling that speedstep thing in the kernel. From what I have read (depending on other configuration) this can be a bit too agressive and change the freq too much and too often, resulting in lag
_________________
#define HelloWorld int
#define Int main()
#define Return printf
#define Print return
#include <stdio>
HelloWorld Int {
Return("Hello, world!\n");
Print 0;

[Anon-E-moose]

Running an AMD 64 x2 and from my view 2.6.35 is faster than 2.6.34 both on boot up and normal running, X is noticeably faster.
That's using the zen kernels with BFS instead of CFS, don't know if that makes a difference.
I have never used the cpu freq settings (turned off in bios), so I can't say it that's an issue.
_________________
UM780, 6.12 zen kernel, gcc 13, openrc, wayland

[drescherjm]

[drescherjm]

I was wrong about the typing being fixed in the pappy seeds kernel. It still is ridiculously slow if I ssh in.

I tested disk performance for the boot hard drive (2TB WDC green) and that looks fine (well at leas sequential performance).

[s4e8]

Disable CONFIG_INTEL_IDLE or add intel_idle.max_cstate=0 to boot cmdline.

[drescherjm]

I am pretty sure that is disabled although since I am at work and the box if off at home (its 90F+ / 32C+ here this week) I will not be able to check for at least 8 hours..
_________________
John

My gentoo overlay
Instructons for overlay

[Anon-E-moose]

Not sure what motherboard/chipsets you are using, but have you googled your motherboard/chipsets and 2.6.35 to see if others are having problems?
_________________
UM780, 6.12 zen kernel, gcc 13, openrc, wayland

[drescherjm]

No I have not. There is only 1 chipset for socket 1366 desktop processors though (X58). I can look into that.

http://en.wikipedia.org/wiki/Intel_X58


My board is an EVGA E758-TR 3-Way SLI (x16/x16/x8) LGA 1366
http://www.newegg.com/Product/Product.aspx?Item=N82E16813188046

I will look into that when I get home.
_________________
John

My gentoo overlay
Instructons for overlay

[drescherjm]

[transpetaflops]

I notice you have the now infamous WD Green drives. Please check the SMART attribute Load_Cycle_Count and compare it to the drive's power on hours. The IntelliPark feature of these drives do not play nicely with Linux unfortunately and I experienced the exact same behaviour as you until I threw them out. IntelliPark offloads the drive heads after 8s idle and it takes about 30s to return them, resulting in massive iowait. The desktop drives are spec'd for 300,000 load cycles but it's common to experience tens of them every minute on a Linux system quickly reaching 300,000 and beyond. There is a WD utility called wdidle3.exe that can turn off IntelliPark on some drives or increase the default 8s timeout to over an hour on others. WD doesn't provide it any longer claiming they don't support Linux but it can be found on the net. Google for IntelliPark, wdidle and Linux and it should turn up lots of information for you.

[drescherjm]

Thanks. I thought my drive did not have that because I looked at the load cycles and it was low after the first week but now it is at ~22000 after a month. I have a second 2TB green that has 49 load cycles after over a 1 year. I will make steps to stop the load cycle and reboot on a WDC black drive instead. I am not sure if this explains the situation totally because I do not have the long delays in 2.6.34.
_________________
John

My gentoo overlay
Instructons for overlay

[drescherjm]

I just booted the drive with 5 minute park under 2.6.35-zen and had a 250+ second boot. The interesting thing is to get the wdidle on the drive I used sysrescuecd on a usb stick instead of a dvd+rw to load freedos and the utility. After that I booted from the usb stick to the 64 bit live cd (gentoo based) and there was no slowness under 2.6.35. It seems either my settings, boot device or compiler is causing this. And since in the past I used settings from the sysrescuecd I may rule that out. I have an unused WDC black drive that I can test on. If I get time I will try a fresh stage3 install of 64bit funtoo or one of the other stage3s that are current.
_________________
John

My gentoo overlay
Instructons for overlay

[zx2c4]

I too have this issue with an i7 on 2.6.35 being outrageously slow. Anyone solve it?

[transpetaflops]

Well, if you guys suspect your kernel configuration - try mine. I assume it would be valid on most i7/X58 motherboards with little or no modification.
http://pastebin.com/p9m2dDh2
Double check your HDD controller settings though. My config only contains AHCI, Marvell legacy and MDRAID-1 support.
/Morgan

[zx2c4]

Here's my config, if it helps:

[transpetaflops]

Have you tried a kernel which doesn't contain the BFS/BFQ schedulers you're using right now? They're not part of the vanilla kernel.

[zx2c4]

Yup. I only started using those because I thought it'd fix things.


This is an issue with vanilla.

[transpetaflops]

Your kernel config implies you're running Tuxonice and it probably won't work very well with neither vanilla-sources nor gentoo-sources.

[zx2c4]

tuxonice has nothing to do with this problem. disabling it doesn't change a thing.

again, this is a problem with vanilla

[Anon-E-moose]

http://forum.soft32.com/linux/35-cpu_idle-bug-report-i7-870-cpu-x86_64-ftopict519483.html
_________________
UM780, 6.12 zen kernel, gcc 13, openrc, wayland

[zx2c4]

CONFIG_INTEL_IDLE does not fix the issue -- it's the same with it on and with it off. I think the problem in that linked thread was fixed in a 2.6.35.x patchset.