The patch for closing connections has been commited to the epoll branch. Incidentally, what would it take for epoll to be considered for inclusion into HEAD now? I noticed that kqueue is still missing the two things that were stated that comm_epoll.cc needed (see below):
from comm_kqueue.cc: /* * XXX Currently not implemented / supported by this module XXX * * - delay pools * - deferred reads * * So, its not entirely useful in a production setup since if a read * is meant to be deferred it isn't (we're not even throwing the event * away here). Eventually the rest of the code will be rewritten * so deferred reads aren't required. * -- adrian */ On Mon, 2003-01-27 at 17:22, David Nicklay wrote: > > I finally got a test box for me to work on again. It is a single CPU > 1GHz Pentium III with 2 GB RAM, stock Redhat-7.2 + linux-2.5.59. > Running squid-3.0-devel using the epoll branch and an added patch to > close connections in epoll, I was able to get 1917 requests per second > on i.cnn.net. > > This is a substantial improvement over my previous test on this box with > linux-2.4.9 running squid-2.5.STABLE1. In that test I only got 1333 > requests per second. Unfortunately, there are too many variables for me > to attribute the increase in performance to squid-3, the new kernel, or > the epoll patch. I will try running another test to isolate it if I > can. > > The patch for closing connections in epoll is forthcoming this week. > > -- > David Nicklay > Location: CNN Center - SE0811A > Office: 404-827-2698 Cell: 404-545-6218 -- David Nicklay Location: CNN Center - SE0811A Office: 404-827-2698 Cell: 404-545-6218
