mån 2007-03-26 klockan 11:08 -0400 skrev NGUYEN, KHANH, ATTSI: > I send a load of 4000-5000 request/second to the server. After about 5 > hours, the ability to serve 5000 requests drops and the cache.log is > filled with error 'commBind: Cannot bind socket FD 204 to *:0: (98) > Address already in use You ran out of free ports, all ports occupied by open server connections of TIME_WAIT sockets. TCP is much better at accepting connections than making connections at high rates. Making outgoing connections requires a free port, and the port gets occupied for the lifetime of the connection + TIME_WAIT (2 minutes). The port range used by the kernel is specified in /proc/sys/net/ipv4/ip_local_port_range, and you You can safely use 1024 65535 which gives you 64.5K available ports which should be able to sustain a constant load of 5K connections/s. To go above that you need to either improve how the kernel assigns ports to outgoing connections allowing it to make use of multiple IP addresses, or modify Squid to "manually" assign the local ip:port manually when making outgoing connections. Another thing which significantly improves the situation is to make sure you have support for persistent connections enabled in your backend servers (and Squid). This significantly reduces the amount of connections made by reusing server connections for multiple requests. Regards Henrik
Attachment:
signature.asc
Description: Detta =?ISO-8859-1?Q?=E4r?= en digitalt signerad meddelandedel
