The correct use of fine-grained locking primitives to allow concurrent
execution of key parts of the kernel on multiple processors is a means
to an end, not an end in itself. Having (for instance) per-pool locks
in the pool subsystem doesn't help improve concurrancy if the workload
requires many allocations from a single pool.

With any OS and just about any shared-memory multiprocessor, you can
construct workloads which have linear or near-linear processor
scalability, and workloads with negative scalability (i.e., go slower
as you add more cpu's).

The trick is to allocate your engineering resources to making the
workloads which matter to your users scale well..

						- Bill