Pasi Sjoholm writes:

 > Yeah, when the ksoftirqd is taking all the cpu it will be like that, but 
 > when the kernel is behaving normally the starving diff is between 0->1sec.

 Well ksoftirqd makes your kernel load just visible which is good and 
 ksofirqd gets accounted for this when softirq's get deferred to it.
 It may look like goes from 0 to 100% but thats probably not the case.
 The problem is we can starve userland at high loads. As said we were
 trying some way to cure this I may have some old patch if you like to try.

 Cheers.
						--ro