Allow me to make myself very clear:

- The lexicon data structure is immutable.
- I have named it the ADTDAWG - Adamovsky Direct Tracking Directed Acyclic Word Graph, contained in 4 arrays of basic number types.
- The character set I have chosen is 14 of the best English letters.
- There is a thread that is responsible for words that begin with each of the letters in the character set.
- The threads all call the same recursive word-discovery function, and then modify a set of global time-stamps to eliminate the duplicate word problem.
- They will never try to modify the same time-stamp because they are responsible for a different subset of the lexicon.

- I used mutexes, and condition variables to communicate when work on the current board should begin, and when it has finished.

- My question is this - Do POSIX multi-threads really allow for an optimal implementation of a micro-parallel algorithm? Or am I doing something wrong, because I am only using 45% of the power of my Quad-Core, when I should be maxing it out?

Do you really want to look at the code that I wrote? At this point, I have every reason to believe that it will introduce mass confusion. It is well written, but it requires an in-depth knowledge of lexicon data structure optimization.

So you suggest that each core has a private cache? I would love to see a block diagram of how the Core2 layout works, so that I could stop spinning my wheels on this problem.

A single thread can score 1277 of the best 23 boards found to date, each with a score around 10769 points for the TWL06 lexicon.

That means that the recursive function is being called many, many, many times per second. Should this fact be a concern?