|
I agree that the ~1,000 simultaneous connection count is indeed the
general breaking point for select()- or poll()-based web servers (in my
experience), and epoll() is a much better solution as you say.
Too bad that apache only allows epoll for MPM event models. For the
other interested readers, we're essentially talking about a feature
which is best described here:
http://www.kegel.com/c10k.html
Now, as for the memory pressure mentioned below, I beg to differ a bit ...
But Apache will not handle 1,000 simultaneous connections unless you
have 4GB of RAM, you're on a 32-bit platform, and you have every feature
turned off.
Why is that? I have rarely hit the problems serving 800-1000 concurrent
sessions on 32bit using a normal 2G/2G-split 2.4.x or 2.6.x kernel. I'll
try to get you a serverstats output from one of our customers tomorrow,
if possible.
And then only if you don't want any disk buffer/cache. :)
Again, I believe that if you already hit the memory limits, you did
something wrong in your configuration or setup :).
With typical application server support (e.g., mod_php), Apache will not
reach 1000 processes without something like 8-16G of RAM.
Ohhh, the OP never talked about mod_php or even mod_perl for that
matter. Both are memory hogs, but then again, if you use a proper m:n
threading model, I bet you can still serve a couple of hundred
concurrent connections.
I've never
been able to set MaxClients above 200... Copy-on-write only goes so far.
I would argue that COW kills your performance because your application
was not designed properly :). No pun intended, but I've more than once
fixed rather broken web service architectures based on PHP or Servlets
or JSP or ASP or <insert you favourite web service technology>.
May I ask what your average processing time of a typical deep link
(non-static) pageview is?
Sorry for the tangent, but throttling/DoS prevention is especially
important for any web/application server based on the process model.
DoS prevention does not exist, this topic has been beaten to death
already :). DoS mitigation, maybe yes. Maybe we should define throttling
before continuing discussing its pro/cons. It could very well be that we
agree on that.
Thanks for your feedback and best regards,
Roberto Nibali, ratz
--
echo
'[q]sa[ln0=aln256%Pln256/snlbx]sb3135071790101768542287578439snlbxq' | dc
|
