irq balance problems?

irq balance problems?

[Imre Gergely]

hi

i have kernel 2.6.9, with FC2, on an AMD64 Dual Opteron.

[root@btv 28]# cat /proc/interrupts
           CPU0       CPU1
  0:   22249515   39262513    IO-APIC-edge  timer
  1:          0          8    IO-APIC-edge  i8042
  2:          0          0          XT-PIC  cascade
  8:          0          0    IO-APIC-edge  rtc
 14:          1         13    IO-APIC-edge  ide0
 24:         10  441422068   IO-APIC-level  ioc0, eth1
 25:  171765170       5905   IO-APIC-level  ioc1, eth2
 28:        921     406066   IO-APIC-level  eth0
NMI:      16960       6317
LOC:   61500529   61500464
ERR:          0
MIS:          0

i was wondering, if IRQ28's (eth0) affinity is set to the default "f"

[root@btv 28]# cat /proc/irq/prof_cpu_mask
f
[root@btv 28]# cat /proc/irq/28/smp_affinity
f


and irqbalance is not running, why aren't the interrupts coming from eth0
balanced between the two processors? at least that's what i understood from the
examples in Documentation/IRQ-affinity.txt. are there any other settings/kernel
parameters/compile option one has to set?

(pls msg me in private, i'm not on the list.)
thanks

Re: irq balance problems?

[Arjan van de Ven]

> and irqbalance is not running, why aren't the interrupts coming from eth0
> balanced between the two processors? at least that's what i understood from the
> examples in Documentation/IRQ-affinity.txt. are there any other settings/kernel
> parameters/compile option one has to set?

it'll depend on the chipset. Some round-robin, some don't.
For performance it's better to not round-robin. 

Re: irq balance problems?

[Imre Gergely]

Arjan van de Ven wrote:
>> and irqbalance is not running, why aren't the interrupts coming from eth0
>> balanced between the two processors? at least that's what i understood from the
>> examples in Documentation/IRQ-affinity.txt. are there any other settings/kernel
>> parameters/compile option one has to set?
> 
> it'll depend on the chipset. Some round-robin, some don't.
> For performance it's better to not round-robin. 

is there a way to see for sure? or this behaviour is proof enough that it
doesn't do round-robin?

Re: irq balance problems?

[Arjan van de Ven]

On Tue, 2006-02-21 at 10:03 +0200, Imre Gergely wrote:
> 
> Arjan van de Ven wrote:
> >> and irqbalance is not running, why aren't the interrupts coming from eth0
> >> balanced between the two processors? at least that's what i understood from the
> >> examples in Documentation/IRQ-affinity.txt. are there any other settings/kernel
> >> parameters/compile option one has to set?
> > 
> > it'll depend on the chipset. Some round-robin, some don't.
> > For performance it's better to not round-robin. 
> 
> is there a way to see for sure? or this behaviour is proof enough that it
> doesn't do round-robin?

it's pretty much proof to me already yes ;)

(but why do you want round-robin? it's the worst setting for
performance..... )

Re: irq balance problems?

[Imre Gergely]

Arjan van de Ven wrote:
> On Tue, 2006-02-21 at 10:03 +0200, Imre Gergely wrote:
>> Arjan van de Ven wrote:
>>>> and irqbalance is not running, why aren't the interrupts coming from eth0
>>>> balanced between the two processors? at least that's what i understood from the
>>>> examples in Documentation/IRQ-affinity.txt. are there any other settings/kernel
>>>> parameters/compile option one has to set?
>>> it'll depend on the chipset. Some round-robin, some don't.
>>> For performance it's better to not round-robin. 
>> is there a way to see for sure? or this behaviour is proof enough that it
>> doesn't do round-robin?
> 
> it's pretty much proof to me already yes ;)
> 
> (but why do you want round-robin? it's the worst setting for
> performance..... )

on one of the ethernet cards is much traffic, and there is htb (with many
classes and filters) used. because of this, one CPU is always on 95-100%, the
other is barely used. i wanted to see if it would improve performance if it did
balancing between the CPUs.