[squid-users] 2.4STABLE1 hangs for a few minutes on Linux 2.2.18

From: Alain Fauconnet <alain@dont-contact.us>
Date: Thu, 22 Nov 2001 14:59:03 +0700

Hello -

I've been running a Squid 2.4STABLE1 for quite a few months on
a high-end IBM eSeries servers with ample memory (512Mb) and Ultra160
SCSI disks, with Linux 2.2.18 as the OS (Slackware 7.0)

Apart for a few segmentation violations lately, it's been fairly
stable *EXCEPT* for a recurrent problem: from once a week to several
times per day, it would "stop listening" to its port (3128) for a
length of time ranging from 1 to 15 minutes. It means that a telnet to
this port does *NOT* get connection refused, but the connection
doesn't establish either. It stops to "Trying...". Tcpdump shows the
SYN packet coming, then nothing else.

There is absolutely *NOTHING* in the log files to give me a clue as to
what happens. Neither Squid logs, nor Linux system logs. Things then
resume automagically after a while... until next time. Other network
services (Apache, sshd) on the same box are unaffected. No sign of DNS
resolution trouble at this time.

The server runs rather cool as I can see, no sign of memory or CPU
cycles shortage. Disks are reasonably busy.

I've tried everything I could think of:

- raised the open FD limit to 8192 (and yes, I checked that Squid does
see the new limit and uses them - it shows in the log files). There is
no sign of running out of FDs.
- set the /proc/sys/net/ipv4/tcp_max_syn_backlog to 128 in case the
SYN queue overflows (but no evidence that it did)

Nada. It keeps happening. I have failed to match this with any
specific traffic pattern or anthing else so far.

If anyone can shed some light on this, I'd be really grateful.
Configuration:

Server: Slackware 7.0, kernel 2.2.18, libc 2.2.1
Squid: 2.4 STABLE1, cache_mem 64Mb, cache_disk aufs, 8192Mb

For completeness, this server is also used as a VPN server so it has a
high number of ppp interfaces, but Squid is bound only to the Ethernet
interfaces in the config file.
It's also member of a two-node child-parent hierarchy (as the child).

Thanks for reading,
Greets,
_Alain_
Received on Thu Nov 22 2001 - 00:59:09 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 17:04:25 MST