Craig,

> 	I still maintain, though, that the central issue is how 
> to avoid an endless loop of zero-length writes, and that 
> using a blocking select() in a separate thread is the way to do it.

That point is very well taken. If you look at the description in one of my
recent emails and the new implementation of the socket primitives, it is
exactly what's going on now. In fact, the real fun part about using select()
in a thread is that its got a separate timeout and can be used independently
from the actual blocking behavior of sockets (and because of this it is
possible to design a cleanup procedure without having to terminate the
threads from the outside). About differences in speed - I'm not sure what
difference will be with the overhead introduced by using the current socket
primitives but I'll happily try any benchmark you come up with.

Which reminds me: Do the PWS guys have any benchmarks for testing real live
behavior of the socket implementations?! I'd love to run a bunch of actual
benchmarks just so that one can get a feeling for the difference across
different implementations/platforms.

  Andreas