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
signature.asc
Description: This is a digitally signed message part
