Joe Duffy is building a custom threadpool on his blog, exploring the different trade-offs:

• Building a custom thread pool (series, part 1)
• Building a custom thread pool (series, part 2): a work stealing queue

The first threadpool he designs in part 1 is a very simple one, using a naive algorithme. The second one in part 2 is more interesting, since this allows threads to 'steal' work from each other and tries to take advantage of work still being in the memory cache by using a LIFO (Last In First Out) algorithm.

Update:

Part 3 was posted some time ago, which integrates the queue from part 2 in the pool from part 1.

• Building a custom thread pool (series, part 3): incorporating work stealing queues