[squid-users] Performance problems

Hello,

For a while, we've been having performance problems on one of our proxies.
So far it looks like the machine is responding horridly when memory is
freed.

Here's some sample output from vmstat:

20060717-120000 2 3 0 18332 190544 1013220 1 1 0 3
 0 0 2 2 1 1
20060717-120100 2 0 0 20744 191040 1014308 1 1 0 3
 0 0 2 2 1 1
20060717-120200 2 0 0 20620 191576 1013444 1 1 0 3
 0 0 2 2 1 1
20060717-120300 2 0 0 20828 192012 1012816 1 1 0 3
 0 0 2 2 1 1
20060717-120400 1 0 0 29832 192392 1002868 1 1 0 3
 0 0 2 2 1 1
20060717-120500 1 0 0 58108 192524 971760 1 1 0 3
0 0 2 2 1 1
20060717-120600 2 0 0 69172 192784 965056 1 1 0 3
0 0 2 2 1 1
20060717-120700 2 0 0 45644 193200 988184 1 1 0 3
0 0 2 2 1 1
20060717-120800 2 0 0 24668 193604 1008776 1 1 0 3
 0 0 2 2 1 1
20060717-120900 2 0 0 21576 194048 1011436 1 1 0 3
 0 0 2 2 1 1
20060717-121001 4 0 0 18056 194400 1010904 1 1 0 3
 0 0 2 2 1 1
20060717-121100 2 0 0 18652 194904 1013504 1 1 0 3
 0 0 2 2 1 1

Between 12:04 and 12:07, the machine was responding very poorly.

Output of 'free':

             total used free shared buffers cached
Mem: 2055448 2034576 20872 0 202756 1002868
-/+ buffers/cache: 828952 1226496
Swap: 8388600 0 8388600

Specs of the machine:

Dual processor Intel(R) Xeon(TM) CPU 3.20GHz, 2 GB memory, machine has 3
HD's: 2x72.8GB mirror for OS/software/logs and 1x72.8G single disk for
cache.

OS: Linux kslh086 2.4.21-37.ELsmp #1 SMP Wed Sep 7 13:28:55 EDT 2005 i686
i686 i386 GNU/Linux (upgrade to 2.6 is unfortunately not possible)

Disks are all ext3.

Relevant squid.conf lines:

cache_replacement_policy heap GDSF
memory_replacement_policy heap GDSF
memory_pools off
cache_swap_low 90
cache_swap_high 95
maximum_object_size 64 KB
maximum_object_size_in_memory 8 KB

Apart from squid, Apache httpd 2.2.2 and BIND 9.3.2 (as a caching DNS
server) are running on this machine.

Relevant Apache config:

MinSpareServers 5
MaxSpareServers 5
StartServers 5
MaxRequestsPerChild 0
MaxClients 5

I've already minimised the cache (it's only 1G large now) to see if the
problem was too much disk access, but no luck.

Normal cache usage is about 300-350 req/s, throughput is about 2.5MB/s,
and usually there are about 2000-2500 fd's open (proxy is configured to
run with 8192 available fd's, we can't lower this as the peak usage seems
to be about 6000 fds)

Anyone has any ideas what might cause this? Directions to search for?

Joost
Received on Mon Jul 17 2006 - 04:33:57 MDT