[squid-users] bad performance

hi out there,

we are trying to solve some performance problems with our squid-proxies.

--- cut ---

first our configuration:

at the moment there are at about 10.000 users in 150 co-locations. every
co-location has it's own squid-proxy. the topology looks like:

        co-location- => central proxy => firewall => DMZ-proxy =>
internet
        user-proxy
           
the central proxy is a double-cpu server PIII 1GHz, 1 GB RAM, 18 GB HD,
running suse linux 8.0. we have configured three different proxies on the
central proxy (one intranet, extranet and internet)
we only use parent-proxy-configuration. no siblings, no cache-neighbourhood.

at peak we have at about 100-150 Requests/s with 13-20 MB transfer volume/s.

time transfer volume (MB) Objects/Requests
11.20 13.8325986862183 5691
11.21 18.0507574081421 6150
11.22 14.2347249984741 5348
11.23 30.8025789260864 5866
11.24 13.9124889373779 4969
11.25 13.1664276123047 5122
11.26 13.3603639602661 5872
11.27 13.8448610305786 5930
11.28 19.539514541626 6692
11.29 15.2315979003906 6403

hardware for central proxy:

HP NetServer LPR2000
double cpu, 2*PIII 933MHz, 1GB RAM, 2*18 GB SCSI HD Raid 1 (gives 18 GB to
use)
HD configuration is not the best, I now, but at the moment we can hardly
change
(add further HDs) anything because of closed budgets.
Suse Linux 7.3, Kernel 2.4.10-64GB-SMP
Squid 2.4.STABLE7, configured with
./configure --enable-auth-modules="MSNT" \
             --with-aio-threads=32 \
             --with-pthreads \
             --enable-storeio=null,ufs,aufs \
             --disable-ident-lookups \
             --prefix=/users/www/squid \
             --enable-err-language=German \
             --enable-underscores \
             --disable-internal-dns \
             --enable-gnurege

we are using aufs as cache_dir with 5 GB space.

--- cut ---

we have a 4MBit connection to the internet.
we always believed that our user-proxy is slow (a HP d220, HP-UX 11).
we have done some performance tests and they showed that this HP-UX machine
really was very slow as soon as there were a few connections (5 clients
paralell) .
so we installed a new pc, with athlon 1,4GHz, 512 MB, 20 GB HD
(unfortunaltly only IDE because our budgets are closed at the moment), Suse
Linux 8.0, Kernel 2.4.18-4GB.

our proxies don't use ICP or multicast - as far as i understood this only
was needed with a neighbourhood/sibling-proxy, is this right?

we now thought that this new hardware performs much better, but with only 2
or 3 users surfing over this new proxy the perfomance is bad. even our
"so-believed-slow" old proxy performs much better. one test we have done was
to use a browser on this new hardware and
        a) using proxy-dmz
        b) using itself (so browser and proxy on the same computer)
as proxy. case a) was very fast, case b) was sloooow.

we are now trying to find any information that help us to solve this
problems.
like:
- is the hardware of our central-proxy capable enough to handle those 150
requests/s?

our recent preformance test showed that influence of caching on/off, acls
on/off
seems not to be that great.

Markus Rietzler
* <rietzler_software/>
* RZF NRW
* Tel: 0211.4572-130
Received on Fri Jul 26 2002 - 06:20:18 MDT