[squid-users] Fwd: Squid configuration advise

[Jean Christophe Ventura]

Hi,

I'm currently working to migrate RHEL5 2.7 Squid to RHEL7 3.3.

I have migrated the config files to be 3.3 compliant (CIDR, remove of
deprecated function,change cache from UFS to AUFS) without any change
(cache mem, policy, smp)

The new platform is a 4 node R610 (24 proc hyperthreading activate)
with 48GB of RAM, only 143GB disk in RAID for OS and cache. Each node
is connected to the network using 2x1Gbit bonding 2/3 level (some
network port are available on the server).

bandwidth allocated for Internet users 400Mbit

The difference between the old plateform and the new one doesn't seem
to be very fantastic :P
I have read the mailing list history alot.

Squid release:
So i know 3.3 isn't anymore maintain but this infrastructure will be
not maintain by myself and i don't think that people behind will do
the update them self
If a official repository exist, maybe this question will be reopen
(from what i have read it's more some of you build packages from
source and give them to people)

Squid auth:
It's transparent/basic auth only filtering some ip with acl.

Squid bandwidth:
Currently a squid node treat something like 30/50Mbit (information
recovered using iftop)
>From previous viewed mail i think it's normal for a non-smp configuration

Squid measure:
[root at xxxx ~]# squidclient mgr:5min | grep 'client_http.requests'
client_http.requests = 233.206612/sec
other info
Cache information for squid:
        Hits as % of all requests:      5min: 6.8%, 60min: 7.1%
        Hits as % of bytes sent:        5min: 4.7%, 60min: 4.4%
        Memory hits as % of hit requests:       5min: 21.4%, 60min: 21.5%
        Disk hits as % of hit requests: 5min: 34.7%, 60min: 30.8%
        Storage Swap size:      9573016 KB
        Storage Swap capacity:  91.3% used,  8.7% free
        Storage Mem size:       519352 KB
        Storage Mem capacity:   99.1% used,  0.9% free
        Mean Object Size:       47.71 KB

Now question and advise :

This metrics seem too low for me. anyone of you agree ?

4 node x 50Mbit node= 200Mbit
To treat the maxbandwidth (400Mbit) + the lost of one host i need to
configure 4 thread by node.
Is there any reason or brillant idea for more (i will have some core
still available) ? calculation too empirical ?

This url http://wiki.squid-cache.org/ConfigExamples/SmpCarpCluster
seem to be a good start :P
Using this method i can interconnect each proxy to share their cache
(maybe using dedicated network port). Usefull or not ? may this
increase the hit ratio ? if this idea is'nt stupid interconnet using
the frontend only or directy to each ?

For now i have :
- 100GB of disk available for cache
- 40GB   of RAM (let 8 for OS + squid disk cache related ram usage)

1 front with the RAM cache and 4 back with disk cache.
AUFS or ROCK cache? mix of them ? 50% each ? maybe another rules ?
(i think it's will be linked to the cache content but any advise or
method is welcome)

I can get more speed and/or space for disk cache using SAN, do you
know if the data is sequential or random ?

Any advise/rules to increase the hit ratio ? :)
Any general advise/rules ?

Thanks for your help


Jean Christophe VENTURA