coev vs. Tornado, greenlets

[Yang Zhang]

Do you have any rough performance comparison numbers against
Tornado/Twisted? I saw on the wiki that coev isn't as fast, but I'm
wondering how dramatic the difference is, and whether there's
something fundamentally impeding performance, or if engineering should
be able to remedy things.

I'm wondering how the scalability compares as well. Typically,
asynchronous IO systems really only start shining when there are large
numbers of low-activity sockets, such as in COMET/WebSocket web
applications, due to lower memory pressure. Do you know how scalable
coev is with respect to number of connections, given each connection
has its own full stack, as opposed to the compact heap-allocated state
of explicit event-based programming?

Finally, a question about the inner workings -- coev uses ucontext.h,
but from the wiki it sounds like greenlets are actually copying around
parts of the stack. But does that mean greenlets need to maintain less
than the full C stack per thread, and are thus more compact like
threads? (BTW, is it also the case that Python stack frames are
allocated directly on the C stack?)

Thanks for any answers.
--
Yang Zhang
http://yz.mit.edu/