[squid-users] Large object caching

Ok, I asked this question yesterday and got little response, so I can only
guess that this was the wrong question to ask. Let me ask a different one
instead today.

Can someone tell me where I can find relevant information on proper
configuration of squid for large object caching? Is this even advisable?

I have searched the digests before asking and haven't really found an
answer. Where else should I look?

Best regards,

-mikep

Michael Puckett wrote:

> I am having some difficulty with my squid implementation and have come
> here to the squid experts for your help.
>
> My application is to use squid to manage a small collection (10-12) of
> large (around 2 GB) objects in the cache which don't change at high
> frequency, and deliver these objects to around 40 clients spread across
> 4 network interfaces.
>
> I have a dedicated dual processor Solaris machine with 2 GB of RAM and a
> dedicated 36 GB cache drive and running a quad 100 Mb NIC. The objective
> is to be able to run all 4 ports at switch speed and deliver about 11
> MBytes/sec to each port.
>
> I upped cache_mem to 1500 MB to fully utilize the RAM
> I set store_avg_object_size 1 GB to decrease the structure overhead
>
> I then use wget to fetch a 450 MB object.
>
> This configuration can only deliver an aggregate throughput of about 20
> MB/s, and supports 1 port a full speed, 2 ports at 10 MB/s and 4 ports
> at 5 MB/s. This is less than half of the desired throughput.
>
> I then tried setting maximum_object_size_in_memory to 512MB to get squid
> to retain the object in memory. This worked, as I went from TCP_HIT to
> MEM_HIT, but performance plummeted by about 20X, which was unexpected.
>
> I have also rebuilt with threads and async i/o and run the aufs cache
> type. This did improve performance some, but still not to the point of
> full switch speed. I know the HW is capable of delivering the necessary
> performance as it is possible to repeat the same test against apache
> directly and there is no degradation in performance from 1 to 4 ports.
> It delivers full switch speed to each port.
>
> Can anyone recommend a large object configuration which would work in
> this application? Perhaps a different memory replacement policy too?
>
> Best regards,
>
> -mikep