See Dean's full proposal in process-model.html

----

There is a decent discussion of Win NT completion ports at:

http://premium.microsoft.com/msdn/library/techart/html/servrapp.htm

and a so-so discussion of Win32 multithreading performance at: 

http://premium.microsoft.com/msdn/library/techart/html/threadli.htm

The latter can be a bit trivial (eg. IO bounds tasks generally work better
when concurrently executed in multiple threads, CPU bound tasks work best
when serialized.  duh.) but has some useful stuff in.

It appears that the best way for static content on NT is to use
TransmitFile, completion ports, plus a cache of file handles.  Win32
doesn't support all of that, but that isn't a big deal.

Completion ports appear to be able to map into async IO (ie. aio_* 
stuff) on Unix fine, however from what I understand some systems don't
implement AIO for sockets.