Per-CPU variables

Cross-processor data sharing is expensive
Locking overhead
Cache line bouncing

Often that expense can be avoided
2.6 has a mechanism for per-CPU variables
(See the web site for details)
Examples:
Networking statistics
Scheduler run queues

Things to be aware of
Cross-processor access requires explicit locking
Be aware of preemption 
Use  get_cpu_var(), put_cpu_var()
Total space for per-CPU variables can be limited