/ Zope / Apsis / Pound Mailing List / Archive / 2004 / 2004-01 / Open files, load

[ << ] [ >> ]

[ Typofix and feature add request / Alexander Lazic ... ] [ I'am having problem getting https working on ... ]

Open files, load
"Bo Diechmann" <bo(at)alco.dk>
2004-01-27 12:19:24 [ FULL ]
Hi!
 
We have a website that gets a lot off hits, we currently got 2 backend
servers, and we got a open file problem
backend servers are using about 10% CPU, and nearly no query time on
disk access.
 
when I do netstat -n I get about 3300 connections
but when I do lsof -n around 200.000 open files:
pound      1957    root  394u  sock        0,0            37002100 can't
identify protocol
pound      1957    root  395u  IPv4   36903570                 TCP
10.0.11.125:http->195.249.65.236:1296 (ESTABLISHED)
pound      1957    root  396u  IPv4   39049995                 TCP
10.0.11.125:39700->10.0.11.120:http (ESTABLISHED)
pound      1957    root  397u  IPv4   39049960                 TCP
10.0.11.125:39683->10.0.11.120:http (ESTABLISHED)
pound      1957    root  398u  IPv4   39049751                 TCP
10.0.11.125:39579->10.0.11.121:http (ESTABLISHED)
pound      1957    root  399u  IPv4   39049958                 TCP
10.0.11.125:http->80.208.98.116:4779 (ESTABLISHED)
pound      1957    root  400u  IPv4   39049304                 TCP
10.0.11.125:http->80.164.83.62:1241 (ESTABLISHED)
pound      1957    root  401u  IPv4   39049342                 TCP
10.0.11.125:http->217.157.152.66:37470 (ESTABLISHED)
pound      1957    root  402u  IPv4   39048410                 TCP
10.0.11.125:38927->10.0.11.121:http (ESTABLISHED)
pound      1957    root  403u  IPv4   39049073                 TCP
10.0.11.125:39258->10.0.11.121:http (ESTABLISHED)
pound      1957    root  404u  IPv4   39049964                 TCP
10.0.11.125:http->80.62.178.178:37292 (ESTABLISHED)
pound      1957    root  405u  IPv4   39047823                 TCP
10.0.11.125:http->217.61.189.165:1229 (ESTABLISHED)
pound      1957    root  406u  IPv4   39049427                 TCP
10.0.11.125:http->131.165.108.77:13800 (ESTABLISHED)
pound      1957    root  407u  IPv4   39048989                 TCP
10.0.11.125:http->80.199.112.38:17468 (ESTABLISHED)
pound      1957    root  408u  IPv4   39048609                 TCP
10.0.11.125:39026->10.0.11.120:http (ESTABLISHED)
pound      1957    root  409u  IPv4   39049573                 TCP
10.0.11.125:http->80.163.57.73:1130 (ESTABLISHED)
pound      1957    root  410u  IPv4   39050004                 TCP
10.0.11.125:39705->10.0.11.120:http (ESTABLISHED)
pound      1957    root  411u  IPv4   36967818                 TCP
10.0.11.125:http->195.249.65.236:1455 (ESTABLISHED)
pound      1957    root  412u  IPv4   39047875                 TCP
10.0.11.125:38666->10.0.11.120:http (ESTABLISHED)
pound      1957    root  413u  IPv4   39048610                 TCP
10.0.11.125:http->80.164.149.130:2035 (ESTABLISHED)
pound      1957    root  414u  IPv4   36933092                 TCP
10.0.11.125:http->195.249.65.236:1374 (ESTABLISHED)
 
my question is, how do I get pound to preform better, sa you can see
pound server is
 
10.0.11.125, and I think there is a problem i port numbers, there can
only be 65536 open ports pr. IP number.
 
Can I get pound to work on multiple ip numbers:
 
 
Firewall/iptables -----------------using round robin
------------------------ pound server, with 10 local ip numbers
 
Is this posible.
 
Best Regards
Bo

Re: Open files, load
Robert Segall <roseg(at)apsis.ch>
2004-01-27 13:00:21 [ FULL ]
On Tue, 2004-01-27 at 12:19, Bo Diechmann wrote:[...]

You can have Pound listening on as many IP addresses as you wish, but I
doubt that would help much: if indeed you need that many active
connections you will soon run into the threads limitations. On all
systems I have seen pthreads comes precompiled with a hard limit of 1K
threads, and you need a thread per active connection.

I would however suggest you look into the type of open connections you
have: I suspect a lot of those will show as CLOSING (or FIN_WAIT2 or
similar). Tweak your system to close connections faster and they'll be
gone.

The problem lies with clients which do not close TCP connections
correctly (most notably IE - it likes leaving the connection "half
open"). Having the system time-out them faster is not strictly TCP
compliant but works quite well in practice.

Finally, on your idea of a round-robin: you'll loose session management.
I hope that is acceptable for your application.[...]

MailBoxer