Re: [squid-users] swap growing to big and slowing everything down

Hi Stuart,

On Thu, 13 Mar 2003, Stuart Clark wrote:

> The swap space starts off at 2 meg and slowly grows over 10 hours to 300
> meg. By this time requests are extremely slow with page faults
> occurring.

Sounds nasty, but typical of memory starvation :-(

> Celeron 1.2 GHZ
> 512 RAM
> 1 x scsi 4 GIG (ext3)
> 1 x ATA100 70 GIG (ext3)
> Redhat 7.3
> squid-2.4.STABLE6-6.7.3.rpm
> Swap space 1 gig
>
> The total size of the cache storage was 50 gig but I have reduced it to
> 38 gig to try and fix the problem

Did you check out the FAQ item on typicla memory requirements? I.e. with
38GB cache and your cache_mem setting you should be able to get estimate
of memory requirements.

> my squid box is running about 55 gig of traffic per month (transparent)
>
> The CPU runs at 2% or less
> The ram runs at 495mb constantly without change

That'll be in part because Linux is trying to optimise use of "spare" RAM
(when you have some!) for disk buffers etc.

> Current config settings:
>
> cache_mem 32 MB

Perhaps somewhat high, but obviously doesn't explain 300MB of swap.

> cache_swap_low 40
> cache_swap_high 45
> maximum_object_size 5000 KB
> maximum_object_size_in_memory 8 KB
> cache_dir ufs /cache 4800 16 256
> cache_dir ufs /cache3 73000 16 256
> httpd_accel_host virtual
> httpd_accel_port 80
> httpd_accel_with_proxy on
> httpd_accel_uses_host_header on
>
> What I have tried:
>
> Reducing cache_mem from 64 to 32
> Reducing cache_swap_low from 90 to 40
> Reducing cache_swap_high from 95 to 45
> Reducing reducing maximum_object_size from 10000 KB to 5000 KB
> Reducing maximum_object_size_in_memory from 64 KB to 8 KB
>
>
> Have I got a memory leak?
> If so how can I fix this problem?
>
> IPCHAINS
> Only port 3128 is allowed from my dialup clients ip address pool

What kernel version are your running?

Assuming 2.4.*, ipchains probably isn't your best option (and IIRC, there
have been memory leaks in this department before - anyone confirm this?).
Have you tried running iptables instead of ipchains?

Do you get any relief by re-starting squid? No, I'm not suggesting that
as a solution, just a diagnostic.

Any chance of a snapshot of /proc/meminfo from:
(1) just after a bootup
(2) when the machine is crawling
(3) after the machine has been crawling and squid restarted

Another point to watch is that you'll probably see memory buffers decline
somewhat before getting into serious swap usage - if so, this will already
be hurting performance.

HTH,
Neale.
Received on Wed Mar 12 2003 - 16:57:16 MST