FWIW, I've seen Squid running at 12K req/sec (very small responses, 100% hits from memory). Higher using the multiple instance perl script (it scales reasonably linearly). For a pure proxy (no caching), I'd estimate you could do about 5K req/sec for small responses on modern hardware, based on tests I've done previously. Again, that's one core. Regarding TCP tuning on Linux, see: http://fasterdata.es.net/TCP-tuning/linux.html http://www.psc.edu/networking/projects/tcptune/#Linux Cheers, On 05/02/2010, at 3:56 AM, Kinkie wrote: > On Thu, Feb 4, 2010 at 5:52 PM, Markus Meyer <markus.meyer@xxxxxxxx> wrote: > >> Nice one. I think I can get to testing it next week. But the numbers I >> get out of it must be handled with care. Since this is a pure test >> environment. It's like a "best-case" scenario which sadly never will >> happen in a production environment ;) > > It's ok, and the fact will be highlighted when publishing the results. > It is however common practice of all commercial vendors to use > pure-lab-environment numbers when pitching their offers, and I find it > only fair that we match that with numbers of our own. Also, it'll be a > nice ego-boost for all the Squid community to be able to claim > impressive numbers - hell, your numbers are impressive already.. > > Notice: if you implement multi-instance squid, an added boost might > come from tying each instance to a specific CPU core (on Linux it's > done via the taskset command) > > -- > /kinkie -- Mark Nottingham mnot@xxxxxxxxxxxxx
