On 11/10/11 23:38, mc8647tv wrote:
Good morning,
I just received my new 3 servers for the squid infrastructure.
Each server has:
2 CPU X5690 (each has 6 cores/12 threads, 3.46 Ghz)
12 GB ram (I have kits to upgrade to 24 if needed)
8 hd 146GB 15k rpm (I also have 300/600/900 GB 10k available)
Actual setup are a couple of 2.7 that handles about 2500 clients, balanced via wpad. Squid sends traffic to antivirus software using a standard upstream connection. Internet connection is about 200 mbit.
Now my questions: with all these cores I think that using some form of multi-instance is the best way to use them. My idea was to set 2 disks in raid 1 for boot and log and dedicate each other disk to one squid instance, 30 GB for coss and 100 GB for aufs.
Sounds good to me.
But I also got plenty of ram, 12 or 24 GB... so I also had the idea of a ram only squid...
If you want to go the way of multi-instance you get to choose between
any number of combinations and connection topologies.
The simplest ones are:
* 2-layers with a gateway instance (RAM cache only) using CARP to
balance between several backend instances doing disk caching.
* 1-layer with all instances receiving in parallel. Possibly with
HTCP, ICP or multicast ICP over localhost to reduce the file duplication
overheads.
* mix of the above. Several front-end instances balancing between
backends.
* more exotic setups with other balancing algorithms than CARP.
I can't really say which is the fastest. We don't have any good solid
benchmarking results for modern Squid yet.
Either way the AV will continue to be the bottleneck. So probably aim
for the CARP or topologies to maximize the de-duplication of stored
objects. That way you minimize the AV delayed MISS traffic.
What would you suggest ? We are open to any suggestion, from s.o. to configuration...
Thanks
Francesco
Amos
--
Please be using
Current Stable Squid 2.7.STABLE9 or 3.1.15
Beta testers wanted for 3.2.0.12
