Curtis <curtis.schofield at gmail.com> wrote:> I have read your tun in article and I have tried to make sysctl
> changes - the benchmarks are not very significant. I love how well the
> signals are laid out and how excellently unicorn responds to them - I
> am hoping to get back to mongrel like performance . We were using
> nginx to haproxy to 32 mongrels
>
> No I have set up a unix domain socket with 2mb send and receive buffer
> and a backlog of 2048 - I didn''t see a change with the send
receive
> buffer tuning and I have yet to try a smaller backlog to see. If that
> helps. Any info?
Hi, I''m not sure what size of requests and responses you get, but 2mb
seems like a lot, especially for requests (unless you handle a lot of
big uploads).
Even with large requests/responses, huge buffers means more pressure on
the kernel allocator and you''re also hurting CPU cache locality. So
larger buffers can hurt performance as well as help, I''ll make a note
of
that in the documentation.
The backlog shouldn''t change performance much. A larger one can let
you
withstand short traffic bursts and increase reliability while
benchmarking.
> I am tuning for osx in dev and centos5 in production. Thank you for
> your great work cleaning mongrel into a first class unix friend.
You''re welcome!
--
Eric Wong