I believe that it compiles and even runs without error on s390 (given that its only interface to the kernel is via /proc).  That being said however, I can't find any reason why its usefull on those arches.  While one can steer interrupts to various cpus in a guest account (either via irqbalance, or manually via /proc/irq/<irq #>/smp_affinity), theres no guarantee that those interrupts will actually be steered to different physical cpus.  Since you're always running on a hypervisor on these arches, theres always going to be a disconnect between what irqblance tries to do and what acutally happens on the metal.  If you can come up with some evidence that irqbalance produces a benefit on these systems consistently and predictably, then I'll happily consider it, but just the fact that its runs without falling over isn't reason enough to enable it.

From what I understood from one of our s390[x] kernel developers, he agrees with Neil: the interrupts in s390 are floating interrupts, that is, is not possible/there is no intention to choose the CPU, so s390x won't benefit form irqbalance.