Per-CPU and per-task/process/thread data.
Function shipping vs. data shipping.
Big question: how much local vs. global processing? How frequent, how expensive, ... Better to divide or to centralize?
Relationship to map/reduce? Message passing!
@@@ populate with problems showing benefits of coupling data ownership
with other approaches.
For example, work-stealing schedulers.
Perhaps also move memory allocation here, though its current location
is quite good.