On Mon, 7 May 2007, Pedro Larroy wrote:

> I'm using curl_multi_socket functions in a project and I've read in the
> "roadmap" for the hiper api that benchmarking needs to be done.

Actually, we're already seeing it run and scale fairly well, but of course
more and better benchmarking will always be appreciated and useful. If nothing
else, at least for the pure "marketing" value!

> I think I might be able to help since I have collected many urls and I have
> been using curl_multi_socket with epoll successfully for some time now
> (thanks to this lists and the lastest patches).
>
> I can provide around 1e5 - 1e6 unique, valid urls to test.

Using various remote URLs is however not the best way of doing benchmarking
since then a lot of the speed and responses etc will depend on all the various
remote sites and networks.

A better test is over a local network with a controlled server with the
particular contents and behaviours you want to test and then run the test from
a machine nearby.

A truly good test should be able to repeat more or less identically so that
you can tweak things and rerun and see if the tweaks made any differences.

I think Robert Iakobashvili's both patches for faster hashes and bigger hash
tables are good candidates for inclusion to boost performance even more.

-- 
  Commercial curl and libcurl Technical Support: http://haxx.se/curl.html