The following test was conducted on a POWER6 box with HZ=250. The isolcpus boot option was used to isolate the CPU being measured from any noise (eg timers, workqueues and user tasks).
I used a slightly modified version of ftq, available here.
The benchmark was run as such:
taskset -c 2 chrt -r 10 ./ftq ...
On this box the timer interrupt takes just over 5us to complete. That means we regularly take a 5% slowdown in any 100us period.
After applying Josh Triplett's patch that extends the concept of NOHZ to userspace:
As before there is some noise as the benchmark starts, but other than that there appears to be only one interruption in the 10 second run.