Re: [squid-users] Performance with DeLL/XEON

Hello Henrik...

First of all thank you very much for beeing so helpful and giving me so
much information. I waited a week to test the new configuration using
some information you and the community gave me.
I disabled the HyperThreading feature of my XEON 1.6 and recompiled my
kernel without SMP. I did recompile my squid too.

Now I'm getting 90 req/s before my CPU is 100% filled. I lowed my
mem_cache to 32MB too.

The box is not running any other services, just squid, it has 32 child
processes and I'm still using aufs.

As you said in your other email, I judged it was stable, in fact it is,
it's not halting anymore and I'm not getting these 'queue congestion'
messages (thanks to your help), but my CPU is always 90%-100% when I get
close to 100 req/s and 1500+ file descriptors. By your experience, is this
normal?

Please, if you prefer to point me some readings, ressources that talk
about these issues, I'll thank you very much.

Sorry again for my poor English, friend.

I'm using this processor + 1GB of RAM to

On Sun, 8 Jun 2003, Henrik Nordstrom wrote:

> On Sat, 7 Jun 2003 macaco@ns1.twi.com.br wrote:
>
> > The hardware we use exclusively for proxy-cache is:
> > DeLL Power Edge 1400 XEON 1.6 (Hyper Threading, so the kernel is compiled
> > SMP)
>
> Squid generally have little or no use of SMP..
>
> > Seagate 10K rpm . 30 GB with reiserfs (notail and noatime) for the cache
> > 1 GB . RAM
>
> Squid really puts a high load on the drives seek times, and you want more
> than one drive for a good Squid performance.
>
> > I compiled SQUID-2.5-STABLE3 using:
> > ./configure --enable-linux-netfilter --disable-ident-lookups
> > --enable-removal-policies=heap,lru --enable-delay-pools
> > --enable-cache-digests --enable-pool --enable-truncate
>
> What is --enable-pool?
>
> > --enable-stacktraces
>
> does not hurt.
>
> > --enable-xmalloc-statistics
>
> this should not be used in production.
>
> > --enable-underscores
> > -with-pthreads --enable-storeio=aufs --enable-async-io=128
>
> 128 threads is a definite overkill and will make your Squid fall over when
> there is too much I/O pending due to the lack of seek kapacity in your
> harddrive..
>
> > --enable-kill-parent-hack --enable-time-hack
>
> Are you absolutely sure you want these?? I would not...
>
> > --enable-default-err-language=Portuguese --with-aufs-threads=10
>
> Hmm.. You specify the number of threads twice.. how many do you actually
> want?
>
> > cache_access_log /var/squid/logs/access.log
> > cache_log none
>
> You want cache.log. Disabling this is just silly unless you are
> configuring a Squid with no disks at all..
>
> > cache_dir aufs /squid/cache1 10000 25 256
> > cache_dir aufs /squid/cache2 10000 25 256
> > cache_dir aufs /squid/cache3 10000 25 256
>
> I assume these are jsut partitions on the same drive. For Squid you should
> use one partition per drive, and more than one drive.
>
> > buffered_logs on
>
> does not make a noticeable difference.
>
> Regards
> Henrik
>
>
Received on Thu Jun 12 2003 - 14:26:24 MDT