Re: [squid-users] Tuning Squid for large user base

Hi everyone,

Wanted to post an update. We have not been as active in testing this week
as we found many sites required their upstream servers to know the
originating IPs. I was patching Squid and Linux for the cttproxy method,
but now understand neither is quite ready to test with the latest Squid
stable and Linux 2.6. Other information follows:

On Sun, 7 Mar 2004, Henrik Nordstrom wrote:

> On Sun, 7 Mar 2004, James MacLean wrote:
> > 6MBytes. 620+ sites. Thousands of client computers :).
> Approx how many of those client computers are actively surfing at a given
> point in time?
> I think you should split this load on multiple Squids. You already
> indicated you have a SMP machine, in such case running more than one Squid
> on the same server is possible.

Broke out 4 squids on the one computer and divided the traffic of our B
class over them. Each Squid got a 3G disk cache and 64M cache_mem. From
everyone's feedback, some settings adjusted were :

fqdncache_size 16536
maximum_object_size_in_memory 8 KB
request_header_max_size 1000 KB
half_closed_clients off
quick_abort_min 0 KB
quick_abort_max 0 KB

All in an Intel(R) Xeon(TM) CPU 2.40GHz X 2 (hyperthreading) PC with 1G
or RAM (more being purchased :)).

This time when the squids were brought online, all cpus started to show
quite a bit of activity. The PC, which had always remained responsive on
previous squid tests, began to labor. Obviously it was working much harder
than having only one squid running :). vmstat showed some blocking, but
nothing very high. No swap activity. squids did not each stick to one
particular CPU, but instead swapped around. Possibly there is the ability
to give them CPU affinity, but I did not investigate that to see if it is
possible.

With all that, the FDs went up to around 3,000 per squid. We only ran it
for approximately 10 minutes and can not say much about the response time
other then it did still appear to slow down. It is very hard to know for
sure because our Internet pipe is always full and without 2 parallel links
to the sites we are testing, this slowness might have nothing to do with
Squid.

We are still preping to do more testing, but wanted to pass in this
information just incase anyone wants to throw out more ideas :).

thanks,
JES

-- 
James B. MacLean        macleajb@ednet.ns.ca
Department of Education 
Nova Scotia, Canada