Very poor IO performance (high CPU load), libata

Very poor IO performance (high CPU load), libata

[Carsten Otto]

[-- Attachment #1: Type: text/plain, Size: 1471 bytes --]

Hello (again)!

System specs below.

My system sometimes[1] crashes in the sense that every HDD access is slow
as hell and most of both CPUs' power is used for IO waiting. Then "top"
shows more than 80% "wa" and the corresponding counter in /proc/stat
increases a lot. When plotting these values with rrdtool I see up to 5%
IO wait with normal behaviour and peaks with up to 40% under heavy load.

In the problem case the IO wait never drops below 40% no matter how light
the HDD load is. Logging in into the system via SSH then takes more than
two minutes and other tasks are equally slow.

Disabling AHCI/ACPI/SMP in the kernel did not help.
I also tried several kernel versions including 2.6.16 to 2.6.18 (rc,
mm).

/proc/interrupts shows a lot of interrupts for my network cards, but
disabling the network does not solve the problem.

A reboot usually solves the problem for some time.

There might be a problem with my hardware (in a not yet determined
device) causing this problem. But as long as I do not know what is wrong
I still see the chance of a software error in the kernel.

System specs:
- Mainboard with Intel 945p and iCH7R (Foxconn 945P7AA-8EKRS2)
- Intel Pentium D 805 (Dual Core, 2.66 GHz, 1 MB cache each)
- 4x Maxtor MaXLine III (300GB, 16MB Cache, SATA2)

[1] Sometimes: More than daily, at the moment I am lucky and noticed no
problem for hours.

Thanks a lot,
-- 
Carsten Otto
c-otto@gmx.de
www.c-otto.de

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

Re: Very poor IO performance (high CPU load), libata

[Carsten Otto]

Problem solved!

On Sat, Jul 15, 2006 at 07:45:57PM +0200, Carsten Otto wrote:
> In the problem case the IO wait never drops below 40% no matter how light
> the HDD load is. Logging in into the system via SSH then takes more than
> two minutes and other tasks are equally slow.

Okay, the wait percentage is no problem, but the speed is.
I had to use a second power supply for two of the four SATA disks,
because the main supply is not good enough and causes crashes and
reboots (see my other threads, I will answer there too).
In a test that checked each disk seperately I noticed that exactly those
disks connected to the second power supply perform very poor with write
accesses (65 MB/sec vs. 2 MB/sec). I replaced the second (bad) power
supply with a very good one and now all disks are working fine. My new
main power supply should arrive soon.

I still can't understand why disks work slower when provided with
"wrong" power... The disks are Maxtor MaXLine III 7V300F0.

> There might be a problem with my hardware (in a not yet determined
> device) causing this problem. But as long as I do not know what is wrong
> I still see the chance of a software error in the kernel.

Yeah, the kernel is okay :)

PS: Now a very annoying hunt for some error in my computer ends after
about three weeks. The system now is stable and fast and I am happy...

Bye and sorry for bugging you,
-- 
Carsten Otto
c-otto@gmx.de
www.c-otto.de