2006-12-07 13:28:55

by Hartmut Manz

Subject: SCSI Controler SCRU32 becomes realy slow on the latest kernels

I am using here 2 machines with the ICP Vortex SCSI-Controler SCRU32 for some
years now.

After switching one machine from Debian 3.1 (sarge, with Kernel 2.6.8) to the
upcoming Debian 4.0 (etch with Kernel 2.6.17 or 2.6.18) I have noticed that
my scsi-devices become very slow while there is also a dramatical increase in
the sys time needed for I/O operations.

The machine is a dual processor Xeon system (2 * 2.2 GHz) with 2 GB memory.
The used scsi drives are seagate R10k and R15k disks.

i.e. reading 1 GB from the first SCSI drive takes about:
xen-1:/var/log# time dd if=/dev/sda bs=256k count=4000 of=/dev/null
4000+0 records in
4000+0 records out
1048576000 bytes (1.0 GB) copied, 55.3079 seconds, 19.0 MB/s

real 0m55.361s
user 0m0.368s
sys 0m49.107s

There are two strange things:
1. The transfer rate is only 19 MB/sec what is very low for this machine.
The expected value would be at least 50 MB/sec.

2. the system time is in the same range as the real time and thats
realy annoying, on the old kernel the system time for such an operatin
was about 7 sec.

Thanks for any help

Hartmut

--
-----------------------------------------------------------------------------
Hartmut Manz WWW: http://www.intes.de
INTES GmbH Phone: +49-711-78499-29
Schulze-Delitzsch-Str. 16 Fax: +49-711-78499-10
D-70565 Stuttgart E-mail: [email protected]
Gott spricht: Ich lasse dich nicht fallen und verlasse dich nicht.
------------------------------------------------------ Josua 1, 5b ---------

2006-12-10 06:36:34

by Andrew Morton

Subject: Re: SCSI Controler SCRU32 becomes realy slow on the latest kernels

> On Thu, 7 Dec 2006 14:28:44 +0100 Hartmut Manz <[email protected]> wrote:
> I am using here 2 machines with the ICP Vortex SCSI-Controler SCRU32 for some
> years now.
>
> After switching one machine from Debian 3.1 (sarge, with Kernel 2.6.8) to the
> upcoming Debian 4.0 (etch with Kernel 2.6.17 or 2.6.18) I have noticed that
> my scsi-devices become very slow while there is also a dramatical increase in
> the sys time needed for I/O operations.
>
> The machine is a dual processor Xeon system (2 * 2.2 GHz) with 2 GB memory.
> The used scsi drives are seagate R10k and R15k disks.
>
> i.e. reading 1 GB from the first SCSI drive takes about:
> xen-1:/var/log# time dd if=/dev/sda bs=256k count=4000 of=/dev/null
> 4000+0 records in
> 4000+0 records out
> 1048576000 bytes (1.0 GB) copied, 55.3079 seconds, 19.0 MB/s
>
> real 0m55.361s
> user 0m0.368s
> sys 0m49.107s

ouch.

> There are two strange things:
> 1. The transfer rate is only 19 MB/sec what is very low for this machine.
> The expected value would be at least 50 MB/sec.
>
> 2. the system time is in the same range as the real time and thats
> realy annoying, on the old kernel the system time for such an operatin
> was about 7 sec.
>

It would help a lot of we can determine where the kernel is spending all this
time. Can you please generate a kernel profile?

Start a large IO operation then, while it is running, run

#!/bin/sh
sudo opcontrol --stop
sudo opcontrol --shutdown
sudo rm -rf /var/lib/oprofile
sudo opcontrol --vmlinux=/boot/vmlinux-$(uname -r)
sudo opcontrol --start-daemon
sudo opcontrol --start
sleep 10
sudo opcontrol --stop
sudo opcontrol --shutdown
sudo opreport -l /boot/vmlinux-$(uname -r) | head -50

(might need some adjustments for distro variation)

Thanks.