How to avoid interrupt sharing ?

Hello group,

I have a linux used as a router, with 6 ethernet interfaces (two built-in
and four on a PCI card).

When booting Kernel 2.6.14, I see that the two built-in ethernet
interfaces share a single IRQ with one of the PCI ethernet. I.e :
router:~# cat /proc/interrupts
CPU0 CPU1
....
16: 826072312 1 IO-APIC-level eth0, eth1, eth2
17: 95530 1 IO-APIC-level eth4
18: 276321831 1 IO-APIC-level eth3
19: 294 1 IO-APIC-level eth5
....

The machine is an HP DL140 with the two build-in ethernet being Broadcom
ones (driver is tg3) and the four PCI Intel ones (driver is e1000).

Any idea of how to avoid this interrupt sharing ?

I looked the various BIOS options, but there doesn't seem to be anything
of use there.

--
Jean-Francois "Jef" Stenuit, Technology consultant
Welnet - a service of Legend Software SPRL

Jean-François Stenuit wrote:
>
> Hello group,
>
> I have a linux used as a router, with 6 ethernet interfaces (two
> built-in and four on a PCI card).
>
> When booting Kernel 2.6.14, I see that the two built-in ethernet
> interfaces share a single IRQ with one of the PCI ethernet. I.e :
> router:~# cat /proc/interrupts
> CPU0 CPU1
> ...
> 16: 826072312 1 IO-APIC-level eth0, eth1, eth2
> 17: 95530 1 IO-APIC-level eth4
> 18: 276321831 1 IO-APIC-level eth3
> 19: 294 1 IO-APIC-level eth5
> ...
>
> The machine is an HP DL140 with the two build-in ethernet being Broadcom
> ones (driver is tg3) and the four PCI Intel ones (driver is e1000).
>
> Any idea of how to avoid this interrupt sharing ?
>
> I looked the various BIOS options, but there doesn't seem to be anything
> of use there.
>
> --
> Jean-Francois "Jef" Stenuit, Technology consultant
> Welnet - a service of Legend Software SPRL

Hi,

Is your acpi turn on? I have the same problem before and now too!! I
fixed by turn on new interrupt method introduce on new motherboards.

--
jsuthan
Zues linux team
http://www.mypulau.com

In comp.os.linux.networking Jean-Fran?ois Stenuit <(E-Mail Removed)>:

> Hello group,

> I have a linux used as a router, with 6 ethernet interfaces (two built-in
> and four on a PCI card).

> When booting Kernel 2.6.14, I see that the two built-in ethernet
> interfaces share a single IRQ with one of the PCI ethernet. I.e :
> router:~# cat /proc/interrupts
> CPU0 CPU1
> ...
> 16: 826072312 1 IO-APIC-level eth0, eth1, eth2
> 17: 95530 1 IO-APIC-level eth4
> 18: 276321831 1 IO-APIC-level eth3
> 19: 294 1 IO-APIC-level eth5
> ...

> The machine is an HP DL140 with the two build-in ethernet being Broadcom
> ones (driver is tg3) and the four PCI Intel ones (driver is e1000).

> Any idea of how to avoid this interrupt sharing ?

What's the exact problem with interrupt sharing?

--
Michael Heiming (X-PGP-Sig > GPG-Key ID: EDD27B94)
mail: echo (E-Mail Removed) | perl -pe 'y/a-z/n-za-m/'
#bofh excuse 254: Interference from lunar radiation

On Sat, 21 Jan 2006, Michael Heiming wrote:

>
>> Any idea of how to avoid this interrupt sharing ?
>
> What's the exact problem with interrupt sharing?

On multi-processor systems, whith high interrupt rates (i.e. loaded
network interfaces), you can dedicate one processor per interrupt and
therefore improve troughput.

Almost useless on uniprocessors with network load < 100 kpps

--
Jean-Francois "Jef" Stenuit, Technology consultant
Welnet - a service of Legend Software SPRL

In comp.os.linux.networking Jean-Fran?ois Stenuit <(E-Mail Removed)>:
> On Sat, 21 Jan 2006, Michael Heiming wrote:

>>
>>> Any idea of how to avoid this interrupt sharing ?
>>
>> What's the exact problem with interrupt sharing?

> On multi-processor systems, whith high interrupt rates (i.e. loaded
> network interfaces), you can dedicate one processor per interrupt and
> therefore improve troughput.

Ever tested this? From my experience Linux can easily saturate
100 Mbit/fd links no matter if interrupts are shared or not,
using GB lan the problem is usually the storage needed to keep up
with the data.

[..]

--
Michael Heiming (X-PGP-Sig > GPG-Key ID: EDD27B94)
mail: echo (E-Mail Removed) | perl -pe 'y/a-z/n-za-m/'
#bofh excuse 30: positron router malfunction

"Michael Heiming" <michael+(E-Mail Removed)> wrote in message
news:r4cia3-(E-Mail Removed)...

> Ever tested this? From my experience Linux can easily saturate
> 100 Mbit/fd links no matter if interrupts are shared or not,
> using GB lan the problem is usually the storage needed to keep up
> with the data.

There are plenty of high-speed networking applications where storage
isn't an issue. Firewalls, for example.

DS