By far the best proposal I've seen so far.
We've seen many different example methods of how to choose which processes
to kill. Moving the algorithm out of the kernel and into a userland
daemon would provide a means by which each and every one of these methods
could be implemented.
We could implement this simply by extending the capabilities of kerneld.
Naturally, kerneld (or a separate kerneld thread), would have to take
special precautions to make sure it is always runnable and also give
itself a higher priority.
-Sam.