Re: [squid-users] Performance tips for accelerator setup

On Sun, 17 Jan 2010 14:12:48 +0100, Markus Meyer <markus.meyer_at_koeln.de>
wrote:
> Amos Jeffries schrieb:
>>> Markus Meyer wrote:
>>> mem - Heap GSDF: fast and small objects preferred. Better to have more
>>> small objects in memory than to read them from disk.
>>>
>>> disk - Heap LFUDA: fast and more popular pictures, more "hot" content,
>>> is preferred.
>>
>> I'd add a COSS directory for best access to the small items going to
>> disk.
>
> Hmm, currently I use 4x 80 GB AUFS directories. Should I completely
> replace them with COSS. What I could understand from the Wiki one should
> combine them with AUFS for bigger files. But what sizes are

Yes. You _need_ to combine with some other type. COSS for small files.
AUFS for the larger ones.
COSS can't handle the multi-MB or ISO type files since it works in MB
'stripes'. Files have to fit inside a stripe completely. I've seen configs
going well with up to 10MB stripes, the default is 1MB.
A COSS directory can sit nicely on the same drive as an AUFS dir, there is
a *very* marginal improvement from giving COSS its own drive.

> recommendable here for max file size in COSS and what should the overall
> size be?

I'd take a small multiplier (2x - 3x) of your average object size. Rounded
up to nearest power of 2.
Should be somewhere 32K - 128K.

>
>> Also, you may want to use CARP from the parents to siblings. That will
>
> We already have such a kind of balancing. We assign specific URLs to
> specific proxy triples. So that they always get the same requests.
> Nothing fancy but it works ;)
>
>> If you are using ICP between the siblings, try HTCP for better peer
HITS.
>
> I switched from ICP to exchanging the cache digest some time ago. Less
> chatter between the Squid triples and we still have a good peer cache
> ratio. But I might check HTCP. Allthough I fear that overhead of HTCP
> compared to ICP outweighs the better hit rate.

Ah, I overlooked that digest mention. I've not checked it out to confirm
but my instinct tells digest should be faster than HTCP. But you never know
until it's tried really.

NP: HTCP is used like ICP with many background requests. But sends the
full HTTP headers to the peers. This looses out in that the HTTP headers
waste more bandwidth, then gains in that peers can send back body data
immediately with their HIT reply (halving the delay on peer HITS).

Amos
Received on Sun Jan 17 2010 - 21:57:46 MST