On Fri, 5 Dec 2003, Andriy Korud wrote: > Nothing except > comm_accept: FD 12: (53) Software caused connection abort > 1-5 times per second. > I don't think problem is with disks: diskd processes use 1% CPU, while squid > process itself - 99-100%. And as CPU usage grows, visual disks activity > reduces.
I've seen this with large numbers of connections open from clients, even at relatively low request rates (though that was with an accelerator setup and there was essentially no disk activity). Are you using a particularly large number of open connections for your testing? I've found Squid 3 with kqueue to be a lot better when thousands of connections are open at a time; you may wish to try that. > proxy# sysctl -a | grep maxfiles > kern.maxfiles: 4136 > kern.maxfilesperproc: 3722 maxfiles seems a little low if you are expecting lots of connections; you'll probably need to set both maxfiles and nmbclusters a bit higher in your kernel configuration. I don't think that increasing it with sysctl will increase maxsockets as well. Increasing net.inet.tcp.tcbhashsize in /boot/loader.conf might be a good idea, too. If you're testing with a large number of connections coming from a single host you'll probably also need to increase net.inet.ip.portrange.last.
