11.11 + Squid 2.6 STABLE3 = "No buffer space available"

[Alexander Gotlib]

I try to use Squid proxy on my HP D390 server (HP-UX 11.11).

All works fine but sometimes I see errors in cache.log. Like this:

2006/10/12 16:22:43| comm_accept: FD 74: (233) No buffer space available
2006/10/12 16:22:43| httpAccept: FD 74: accept failure: (233) No buffer space available

I think that is necessary to change some kernel variable. But which one? Can anyone help me?

[spex]

Hi,

An overview of Squid cache size options is available here:
http://www.visolve.com/squid/squid24s1/cache_size.php

PCS

[Alexander Gotlib]

I don't think that problem is in squid config.

cache_mem 256 MB
cache_swap_low 90
cache_swap_high 95
maximum_object_size 16384 KB
minimum_object_size 0 KB
maximum_object_size_in_memory 32 KB
cache_replacement_policy heap LFUDA
memory_replacement_policy lru
cache_dir ufs /usr/local/squid/cache 10240 32 64

System: hpd390 Thu Oct 12 18:51:52 2006
Load averages: 0.43, 0.43, 0.38
203 processes: 196 sleeping, 7 running
Cpu states:
LOAD USER NICE SYS IDLE BLOCK SWAIT INTR SSYS
0.43 9.4% 0.0% 10.8% 79.9% 0.0% 0.0% 0.0% 0.0%

Memory: 456532K (427212K) real, 674948K (494512K) virtual, 20236K free Page# 1/
15

TTY PID USERNAME PRI NI SIZE RES STATE TIME %WCPU %CPU COMMAND
? 27051 squid 182 20 398M 389M run 53:12 26.91 26.87 (squid)

[Alexander Gotlib]

I don't think that problem is in squid config.

cache_mem 256 MB
cache_swap_low 90
cache_swap_high 95
maximum_object_size 16384 KB
minimum_object_size 0 KB
maximum_object_size_in_memory 32 KB
cache_replacement_policy heap LFUDA
memory_replacement_policy lru
cache_dir ufs /usr/local/squid/cache 10240 32 64

System: hpd390 Thu Oct 12 18:51:52 2006
Load averages: 0.43, 0.43, 0.38
203 processes: 196 sleeping, 7 running
Cpu states:
LOAD USER NICE SYS IDLE BLOCK SWAIT INTR SSYS
0.43 9.4% 0.0% 10.8% 79.9% 0.0% 0.0% 0.0% 0.0%

Memory: 456532K (427212K) real, 674948K (494512K) virtual, 20236K free Page# 1/
15

TTY PID USERNAME PRI NI SIZE RES STATE TIME %WCPU %CPU COMMAND
? 27051 squid 182 20 398M 389M run 53:12 26.91 26.87 (squid)

[Alexander Gotlib]

As for as I know in FreeBSD and Linux my problem fix by increase kernel variable "mbufs".

But in HP-UX kernel I cant't find nothing like this...

[Denver Osborn]

Could be tcp queue full which causes accept to return ENOBUFS. man accept(2).

check "netstat -p tcp" and look for "connect requests dropped due to full queue". If connections are shown as being dropped, you can bump it up with "ndd -set /dev/tcp tcp_conn_request_max NN". To make it persistent add value to /etc/rc.config.d/nddconf file.

There still a chance that if connections are being dropped due to full tcp queue it isn't because the system has reached it's limit, but the squid proxy's listener has. I'm not sure if squid has any type of config for queuesize, but something you should also note.

Hope this helps,
-denver

[Denver Osborn]

on second thought... :)

The squid accept() ENOBUFS probably isn't tcp_conn_request_max related, but it couldn't hurt to check "netstat -p tcp" output for connect requests dropped.

Check out Rick Jones' reply on this thread.
http://forums1.itrc.hp.com/service/forums/questionanswer.do?threadId=237561

-denver

[Alexander Gotlib]

> check "netstat -p tcp" and look for "connect
> requests dropped due to full queue".

:-( Nothing wrong...

# netstat -p tcp
...
0 connect requests dropped due to full queue

rick jones wrote:
> The only "fix" is to make sure the server
> application can get around to calling
> accept() in time - either by making the
> server application run faster, or by making
> the client more forgiving.

Hmmm... And what can I try to do with my Squid?