Poll-splitting

Okay, now _this_ is a little odd....Further to my email about splitting
poll() structures into groups according to the likelyhood of getting a
ready-response, I thought..well..hey, tcp connections are buffered in
the kernel. We can probably get _some_ performance improvement by only
calling poll on (say) a third to a quarter of the descriptors per pass,
and letting the kernel buffers take up the slack for the ones we pass
over until a later poll.

Well, my first hack was less than stellar (I think rounding was causing
me to miss a descriptor), so I decided just to call poll with _one_
descriptor of the structure at a time...

Now it seems that I'm getting far better performance out of the damn
thing. Very snappy response on misses. For a low number of users, mind
you. Just me at the moment. Now that sort of makes sense, in a way, and
starts me thinking about windowing the poll size. Hmm.

D
Received on Fri Nov 26 1999 - 17:57:19 MST