performance problems again

From: Javier Puche. CSIC RedIRIS <javier.puche@dont-contact.us>
Date: Fri, 26 Feb 1999 17:43:15 +0100

Hi all,

 I'm getting into trouble with squid again, maybe someone can help me
again:

1.- the real problem:
 during some performance degration stage I describe later, Storage LRU
Expiration age suddenly decreased from 6 days to 0.10 days and stays
there now that the performance problem dissapeared.

2.- Some questions:
 the problems appear to happen with http requests per minute over 2000
and icp over 6000; I wouldn't say that should me too much for a Sun
Ultra-2 with six 4G ultra-scsi disks (on different array controllers)
for the cache. Might any big-cache manager confirm that is not really a
big load ? do you ever run up to 30 dnsserver on such machines ??

3.- the performance problems stage description:

 after the storeSwapOutFileOpened 'cache5p directories, etc' problems
last week I decided to reinstall clean Solaris 2.7 (bare-bones
installation to avoid any disturbing piece of software) machine with
priority-paging, fastfs, noatime, nolargefiles, tcp-ip tuned, etc.. and
recompiled squid-2.1PATCH2 this time only with the
clientHandleIMSReply-leak patch.

 The problem is that this morning at peak hour I started to see again
cache-hits-median-service-times getting over 1 second for some time.
During that time there were no special signs of things going awry, I
checked all system stats I know and cachemgr. No special squid log
lines neither system console warnings. The only thing was dns lookups
getting too slow and the 30 dnsservers being busy (and as a consecuence
file descriptors dangerously getting close to 4000). I went to 32
dnsservers (the maximum) but I soon realised that the problem was the
poor performance of dnsserver and pointed the machine to a less-loaded
dns server; things got better but the 0.10 expiration age I mentioned
above keeps there :-(

 I've limited the max_open_disk_fds as Duane greatly suggested me last
week to check if it makes things better next time it happens.

Thanks again and have you all a nice weekend,

Javier Puche.

PD: I don't find the link to the nice nlanr SNMP squid stats of all its
caches (not pointed from http://www.ircache.net/Cache/Statistics/) I
saw litle time ago, anyone has it handy? thanks.
Received on Fri Feb 26 1999 - 09:46:42 MST

This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:44:45 MST