Re: 2.5.67-mm4 & IRQ balancing

[Andrew Morton <akpm@digeo.com>]

Philippe Gramoullé <philippe.gramoulle@mmania.com> wrote:
>
> [ SMP IRQ distribution ]
>
> Is this what you are looking for ? and are the values changes meaningful ?

Looks good to me.  But it didn't affect your machine at all, did it?

This stuff only counts when the machine is doing a lot of work.  The current
IRQ balancer works well under high interrupt frequencies, but does quite the
wrong thing if you're doing a lot of softirq work at low interrupt
frequencies (gige routing with NAPI).

My gut feel is that we'll never get this right with a single in-kernel IRQ
balancer.  So the proposal is to pull the IRQ balancer out altogether and to
then merge Arjan's userspace balancer into the main kernel tree.

It's a little radical to go placing userspace daemons into the kernel tree,
but I think it is appropriate - this thing is very tightly coupled to the
kernel.

The proposal has these advantages:

- No version skew problems: if the format of /proc/interrupts changes, we
  patch the irq balance daemon at the same time.

- Can build irqbalanced into the intial initramfs image as part of kernel
  build. (lacking klibc, we would need to statically link against glibc)

- Doing it in userspace means that we can do more things.

  - The balancer can "know about" the differences between NICs, disk
    controllers, etc.

  - The balancer can be controlled by config files: "I am a router"

  - The balancer can support non-x86 architectures


Anyway, that's the theory.  None of it has been done yet.