Maybe there is someone out there to do some benchmarking / testing /
commenting on the changes ?

The following patch adjusts the irq balancing a bit.
1) At least on my 2-CPU smp machine IRQs have been bouncing between CPUs.
This seems to happen because the imbalance-value between the CPUs was
noticed to be high enough for an IRQ-switch to a different CPU. After
switch-over the scenario not really improved -> switch back again.
Maybe the modification helps improving this a bit by implementing a
"look ahead" if moving the IRQ really makes sense.
2) As far as I understand, putting all interrupts on one CPU should be
preferred in case of low interrupt load. (cpu cache etc.)
This patch also adds a fallback-behavior if  the load of interrupts
drops again.
Currently I decided to add a processor variable to irq_desc structure.
I´m not sure if  this is the right way to go but I needed the
information on which CPU an IRQ is currently on. (to decide whether an
IRQ has to be switched back to the "primary CPU")
3) Now '/proc/interrupts' also shows the CPU-no the interrupt currently 
is routed to. (Not sure if this breaks some tools ?!?)