How is the Mathematica kernel structured internally, and how does evaluation work?

Question

There is so much information available for Java--the internal details of (for example) the JVM memory manager are well-documented, and the performance considerations arising from its interaction with user code have been publicly discussed. I am interested to know if there is any comparable documentation for the Mathematica kernel. Specifically:

• What are the major steps involved in evaluation?
• Which of these are the most important from a performance standpoint?

Is there any public documentation available about how all of this is implemented? If there is, I wasn't able to find it after searching the Internet: neither the mechanics of the evaluation process, nor the internal workings of particular functions, seem to be known outside of WRI.

Answer 1

To give references to the first part of your question:

The best description about the main evaluation loop I have ever read can be found in (as already stated in the comments) Power Programming with Mathematica: The Kernel by David Wagner. A free pdf version of this book can be downloaded here. Therein, he refers to the tutorial Mathematica Internals: A Tutorial by David Withoff.

Although, these documents are rather old, I'm pretty sure they are still correct since this topic covers the core evaluation process which hasn't changed.