Re: wish for an advice

On mån, 2007-08-06 at 21:49 +0200, Jiri Kuthan wrote:

> We do epoll, that shall be fine then.

Good. Your performance should then be on level with the number of
message transactions / second (including connection setup/teardown as
psuedo-messages=, and not dependent on the number of concurrent
connections.

If you see a dropoff in message transactions / second as the number of
connections increase then either

a) The TCP/IP stack isn't properly tuned for very many concurrent
connections. May be a socket hash size that needs to be increased or
similar.

b) You have some internal design error in your software, looping over
increasing amount of data as the number of connections increase.

> If we would like to get some lab intelligence, is all we have to do
> install squid and buy a lot of memory, or is there some other black
> magic? Have you tried more than 4 GB?

I only had 1GB in the host when performing this test I think. But I
don't remember exactly.

Used a heavily modified version of idleconn to simulate a large number
of idling clients with low client requirements. Sources at
http://www.henriknordstrom.net/code/ (requires libevent).

Regards
Henrik