Re: [squid-users] Performances: Global rule to size a server

Le mardi 06 septembre 2011 à 16:15 +1200, Amos Jeffries a écrit :
> On Mon, 05 Sep 2011 20:23:02 +0200, David Touzeau wrote:
> > Dear
> >
> > I would like to create a kind of law calculation in order to quickly
> > calculate server performances to store squid...
> >
> > I know there is a lot of parameters that should make e this
> > calculation
> > more complex, but this is just to be generic.
> >
>
> "generic" is not possible. It boils down to what _your_ users are doing
> is different to _my_ users.
>
>
> > For example :
> >
> > I have 150 users :
> > -----------------------------------
> >
> > Memory : 750K per users = 150x750=112 Mb memory + 300Mb for the
> > system =
> > 512 Mb minimal.
> > Hard disk cache : 50Mb/user = 150*50 = 7.5Go minimal stored disk
> > size
> > for cache..
> >
> > Is it make sense ?
>
> Your aim makes sense in a way. The metric of "per user" does not relate
> to resource usage though.
>
> This is solely because one user could be making no requests at all or
> several thousand per second. And they can switch between these
> behaviours and random values in between without notice. I have seen a
> network happily serving 15K users with one Squid on a 50MBit uplink, and
> also a GigE network brought to its knees by just a few users. "Normal"
> web traffic in both cases.
>
>
> The Squid relevant metrics are closer tied to requests per second, or
> the internal Mbps line speed of HTTP requirements.
>
>
> Minimal stored disk size is always zero. Maximal is best at <80% of the
> size of the disk you plug in. Moderated by at most 2^24 objects per
> cache_dir. This is a fixed limit, so performance there is relative to
> your average cacheable object size. Disks are best sized along the
> lines of: 2^24 * avg object size. Overall disk count multiply that out
> by req-per-sec to the time you want things cached by or run out of $$ to
> buy disks.
>
>
> Memory consumption is dominated by the disk index (10-15 MB index / 1
> GB of storage). And by cache_mem storage, which will suck up every spare
> byte you can give it until itself starts to suffer those 2^24 object
> limits. per-User/connection requirements are measured in KB, so these
> days not really worth worrying over.
>
>
> Amos
>

*Many thanks Amos

It is clear....
A kind of simple law should be useful when an IT need to know quickly
which server can be ordered according users number but i'm agree with
your arguments...
Received on Tue Sep 06 2011 - 10:54:58 MDT