Peter Chubb, UNSW
Is a preemptive kernel worth it?

Making the Linux kernel fully preemptive has the promise of reducing
latency for real-time programs, improving interactive response, and
(very important) making mouse movement on the X server smoother, even
while listening to streaming mpg3s.

Preemption support has been added to all the other architectures
except M68k and IA64, so I added it to the 2.5.45 IA64 kernel. 

There was no user-visible performance difference.

How does one measure whether the preemption code is effective?
Approaches currently being considered are:

  * Directly measuring interrupt latency for real-time processes,
    under various kernel-intensive loads

  * `feel' when using a mouse and video-card, again under various
    kernel-intensive loads.
 
  * Whether MPG3 playback `skips' when using a small buffer, with a
    kernel-intensive load.

  * Instrumenting the preemption path to see how often it's being
    taken.

As there are still times when the kernel is not going to be
preemptible (while holding a spinlock, for example) it's unclear as
yet whether preemption support on IA64 is worth the slight throughput
reduction it entails, or even whether it's working at all.